AI innovation concentration and the governance challenge

Lippoldt, Douglas Working Paper AI innovation concentration and the governance challenge CIGI Papers, No. 292 Provided in Cooperation with: Centre for International Governance Innovation (CIGI), Waterloo, Ontario Suggested Citation: Lippoldt, Douglas (2024) : AI innovation concentration and the governance challenge, CIGI Papers, No. 292, Centre for International Governance Innovation (CIGI), Waterloo (Ontario) This Version is available at: https://hdl.handle.net/10419/299989 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ CIGI Papers No. 292 — April 2024 AI Innovation Concentration and the Governance Challenge Douglas Lippoldt CIGI Papers No. 292 — April 2024 AI Innovation Concentration and the Governance Challenge Douglas Lippoldt About CIGI The Centre for International Governance Innovation (CIGI) is an independent, non-partisan think tank whose peer-reviewed research and trusted analysis influence policy makers to innovate. Our global network of multidisciplinary researchers and strategic partnerships provide policy solutions for the digital era with one goal: to improve people’s lives everywhere. Headquartered in Waterloo, Canada, CIGI has received support from the Government of Canada, the Government of Ontario and founder Jim Balsillie. À propos du CIGI Le Centre pour l’innovation dans la gouvernance internationale (CIGI) est un groupe de réflexion indépendant et non partisan dont les recherches évaluées par des pairs et les analyses fiables incitent les décideurs à innover. Grâce à son réseau mondial de chercheurs pluridisciplinaires et de partenariats stratégiques, le CIGI offre des solutions politiques adaptées à l’ère numérique dans le seul but d’améliorer la vie des gens du monde entier. Le CIGI, dont le siège se trouve à Waterloo, au Canada, bénéficie du soutien du gouvernement du Canada, du gouvernement de l’Ontario et de son fondateur, Jim Balsillie. Credits Managing Director of Digital Economy (until February 2024) Robert Fay Director, Program Management Dianna English Program Manager Jenny Thiel Publications Editor Lynn Schellenberg Senior Publications Editor Jennifer Goyder Graphic Designer Sami Chouhdary Copyright © 2024 by the Centre for International Governance Innovation The opinions expressed in this publication are those of the author and do not necessarily reflect the views of the Centre for International Governance Innovation or its Board of Directors. For publications enquiries, please contact [email protected]. The text of this work is licensed under CC BY 4.0. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. For reuse or distribution, please include this copyright notice. This work may contain content (including but not limited to graphics, charts and photographs) used or reproduced under licence or with permission from third parties. Permission to reproduce this content must be obtained from third parties directly. For publication enquiries, please contact [email protected]. Centre for International Governance Innovation and CIGI are registered trademarks. 67 Erb Street West Waterloo, ON, Canada N2L 6C2 www.cigionline.org Table of Contents vi About the Author vii Acronyms and Abbreviations 1 Executive Summary 2 Introduction 3 Literature: Rapid Technological Development, Unsettled Scholarship 9 Leading AI-Intensive Firms: A Few Stylized Facts 19 Regulatory Developments 30 Conclusions 35 Appendix 1: Screens Employed in Selection of Top AI Firms among R&D Leaders 37 Appendix 2: Observations on AI Regulatory Regimes in Selected Sample Countries 45 Works Cited vi CIGI Papers No. 292 — April 2024 • Douglas Lippoldt About the Author Douglas (Doug) Lippoldt is a CIGI senior fellow and an international trade economist based in Claremont, California. He served as chief trade economist at HSBC Global Research from 2014 to 2020. Previously, he served in various roles as a senior economist at the Organisation for Economic Co-operation and Development in Paris, France, during a tenure of 22 years. Doug’s early career included seven years as an international economist with the US Department of Labor. He has published extensively on trade topics as well as on related aspects of economic development, labour market adjustment, innovation and intellectual property. Doug holds a Ph.D. in economics from the Institut d’études politiques de Paris (Sciences Po), an M.A. in international studies from the University of Denver and a B.A. in international studies from Washington College in Maryland. He was a Fulbright Scholar at the University of Cologne, Germany, and a Peace Corps volunteer in Burkina Faso. During 2015–2019, he represented HSBC as the deputy delegate on the Business 20 Trade and Investment Task Force. He is currently a contributing author on a Think20 trade task force team. viiAI Innovation Concentration and the Governance Challenge Acronyms and Abbreviations AI artificial intelligence AIDA Artificial Intelligence and Data Act AIIA Artificial Intelligence Industry Alliance APEC Asia-Pacific Economic Cooperation regional forum B2B business-to-business BIS Bureau of Industry and Security (US) CAC Cyberspace Administration of China capex capital expenditure CPTPP Comprehensive and Progressive Agreement for Trans-Pacific Partnership CUSMA Canada-United StatesMexico Agreement DPA Digital Policy Alert DSIT Department for Science, Innovation and Technology (UK) ESG environmental, social and (corporate) governance FATF Financial Action Task Force FTC Federal Trade Commission (US) G7 Group of Seven G20 Group of Twenty GDPR General Data Protection Regulation ICB Industry Classification Benchmark IFAC-BIAC International Federation of Accountants and Business at OECD IP intellectual property IPRs intellectual property rights LLMs large language models NIST National Institute of Standards and Technology (US) OECD Organisation for Economic Co-operation and Development p.p. percentage point R&D research and development SAG-AFTRA Screen Actors GuildAmerican Federation of Television and Radio Artists UNESCO United Nations Educational, Scientific and Cultural Organization USTR Office of the United States Trade Representative VC venture capital WIPO World Intellectual Property Organization WTO World Trade Organization 7AI Innovation Concentration and the Governance Challenge sharing economy.16 There are many possibilities. Regardless which one, IP is likely be part of the discussion. Getting the IPRs protection tuned to address challenges around a new technology may require a period of collecting experience to identify the key issues, impediments and gaps in the system. Use and adjudication of IPRs based on existing laws may provide essential information for subsequently legislating adjustments in the system. With AI, IPR issues are already coming to the fore. Generative AI is raising copyright issues both with respect to training data employed (which may be subject to copyright protection) and with respect to asserting copyright for products developed without sufficient human input. A recent assessment by the US Congressional Research Service advised that it may be premature to legislate on such matters until the courts have a chance to adjudicate and identify any shortfalls in the legal framework (Zirpoli 2023). (Next up for dispute may be the IPR-adjacent issue of control over one’s own image, voice and actions in AI-generated content.)17 Similar issues may arise with respect to patents, where the notion of human-directed creation may need legal clarification. In trade secrets, challenges are emerging in relation to disclosure of algorithms, restrictions on labour mobility and appropriation of data, among other issues, particularly in light of the high value of R&D in the AI sector. Regulatory processes such as customs controls of source code or registration of models for approvals (for example, already required in China and pending in the United States and the European Union; see Appendix2) may present further challenges for protecting trade secrets. It is unclear whether the current IPR 16 For example, copyright protection might be used by a software originator to maintain certain rights over their open-source software creation while making the code readily and freely available for use by subsequent innovators to build upon or even modify, depending on the terms of an open-source licence. 17 The complexity of generative AI concerns was illustrated in the United States by the contract negotiations between the Screen Actors Guild– American Federation of Television and Radio Artists (SAG-AFTRA) and the Alliance of Motion Picture and Television Producers. The deal, ratified on December 5, 2023, defines AI performance concepts such as an “independently created digital replica...used to portray the performer in scenes they did not actually shoot” and generative AI “used to simulate a performer’s voice, facial expressions, and movements to create entirely new content” (SAG-AFTRA 2023a, 2). Each use of AI is subject to the consent of the subject, clear delineation of the AI content to be created, and appropriate compensation. The contract includes a principle of primacy for human performance in the realm of acting, the holding of regular employer-union meetings on the evolving use of AI, and monitoring of implementation (SAG-AFTRA 2023b). framework will prove adequate or novel challenges will require changes in legislation and regulation. Getting the IPRs system tuning right can make a difference for the diffusion of a technology. Work at the OECD and elsewhere has provided evidence that adequate IPR protection can be conducive to trade and investment in technologyintensive products and to technology transfer via licensing of IP (Cavazos Cepeda and Lippoldt 2010; Branstetter, Foley and Saggi 2010; Lippoldt and Schultz 2014). These effects appear to be strongest with respect to patents, but positive associations have also been found in some cases for copyright, trademarks and trade secrets. Where holders of IP have means of protecting their assets, they are more inclined to transfer control to unaffiliated parties (Park and Lippoldt 2005). And likewise, when foreign firms looking to invest directly in a nation are able to protect their IPRs, they are more likely to transfer recent generation technology along with the knowhow to successfully exploit it (Park and Lippoldt 2014). It is likely that similar tendencies will manifest with at least some elements of AI sector technology. A significant distinction with respect to IP and AI concerns the difference in methods between the open-source LLM development community, which employs an open approach with liberal licensing terms, and the developers building proprietary LLMs, asserting their IPRs and holding back some information that conveys commercial advantage. The former are perhaps able to leverage their technical efforts due to inputs from outside contributors, leading in some cases to advances in model development. The latter may be incentivized in their efforts by better access to investor capital and advice in some cases (although, not all will avail themselves of these options), as well as the prospect of financial gains in the event of commercial success. The open-source versus proprietary divide echoes through the AI ecosystem (CB Insights 2023b), although there is significant cross-fertilization among developers using either approach (Kaye 2022). Open-source tools such as PyTorch, TensorFlow and PaddlePaddle have been developed and released by major AI developers (Meta,18 Google and Baidu, respectively). Such tools have benefited, 18 Note that in September 2022 Meta transferred control of PyTorch to the Linux Foundation (Kaye 2022). 8CIGI Papers No. 292 — April 2024 • Douglas Lippoldt among others, smaller independent application developers, including those collaborating internationally and thereby promoting diffusion of innovation. Hosting services such as GitHub and Hugging Face have enabled developers to share code and data. AI as a service provider may deliver proprietary cloud-based solutions to expedite AI development and deployment for users, while in some cases drawing on opensourced inputs (for example, see Wren 2023). Recent top-performing LLMs have tended to be built using proprietary approaches. However, there is an ongoing exchange with the open-source community, and by some measures the performance of opensource LLMs is catching up (Chen et al. 2023; Bremmer and Suleyman 2023, 30). Giant firms such as Meta (with Llama 2) and Microsoft (with Phi), and academic institutions such as Stanford (with Alpaca) have released pre-trained, open-source models, benefiting indirectly from new insights as outside researchers and developers contribute to advance the technology. This leverages talent from the open-source community and can add new vitality to innovation processes. It may also permit small businesses to participate in the AI economy by tailoring and marketing their own elements built on an open-source AI foundation. Given the substantial original LLM development costs, an open-source approach provides some small firms with a useful point of entry to the AI market. The accelerated pace of development from both approaches has led to AI safety concerns, particularly with respect to frontier model development that may unleash new or poorly understood capabilities (Department for Science, Innovation and Technology [DSIT] 2023). Opensource developments raise particular concerns. Downstream evolution of open models may not be adequately tracked; existing guardrails and built-in mitigation measures may not be sufficient to ensure AI safety (Fathi and MacKinnon 2023; Raskin and Harris 2023). As Ian Bremmer and Mustafa Suleyman (2023, 41) put it in a recent article, “The devolved nature of AI development and core characteristics of the technology, such as open-source proliferation, increase the likelihood that it will be weaponized by cybercriminals, state-sponsored actors, and lone wolves.” As a consequence of such concerns, the United Kingdom convened a safety summit among 29 governments in November 2023. One result was the Bletchley Park Declaration whereby participants agreed to redouble their cooperation on AI safety (for example, through identification of safety risks and building risk-based safety policies) and to engage stakeholders to take responsibility when undertaking advanced model development work (for example, through external pre-release safety testing) (GOV.UK 2023).19 Governance Mardi Witzel and Niraj Bhargava (2023) provide a concise overview of the state of play with respect to AI governance.20 They briefly consider key elements of governance such as national strategies, ethical AI frameworks, current and proposed legal and regulatory instruments (notably in the European Union, the United States and Canada), AI standards (highlighting the International Standards Organization, ISO 42001 standard for AI management systems and the NIST AI Risk Management Framework). The authors’ review makes clear that these various elements each have a role and their development remains a work in progress. With respect to balancing among the various elements of governance, Tom Wheeler (2023, 20) notes, “All modern regulations walk a tightrope between protecting the public interest and promoting innovation and investment. In the AI era, traversing the regulatory tightrope means accepting that different AI applications pose different risks and identifying a plan that pairs the regulation with the risk while avoiding innovation-choking regulatory micromanagement.” Bremmer and Suleyman (2023) make the case for early action on the establishment of a global governance mechanism for AI. They highlight the nature of AI as a general purpose technology that is rapidly evolving and capable of myriad applications and, potentially, self-improvement. At some point, AI will exceed human mental performance in a general manner. This has significant potential for good, such as through its ability to accelerate innovation processes, its potential to enhance 19 In relation to international cooperation on AI safety, the Bletchley Park summit was a timely follow-up to the recent Group of Seven (G7) Hiroshima meeting. The G7 Leaders’ Communiqué, May 20, 2023, paragraph 1, stated: “We are determined to work together and with others to: …advance international discussions on inclusive artificial intelligence (AI) governance and interoperability to achieve our common vision and goal of trustworthy AI, in line with our shared democratic values.” See White House (2023a). 20 The authors (2023) tackle AI governance from a unique environmental, social and (corporate) governance (ESG) perspective, whereby AI could be integrated into ESG reporting requirements. 9AI Innovation Concentration and the Governance Challenge productivity, and its anticipated capacity to assist in responding to global crises such as health or climate emergencies. However, AI will cause dislocation in labour markets and it has the potential for misuse in areas such as disinformation, surveillance or even the creation of autonomous weapons, among other possibilities. It may also exacerbate inequalities. Leading AI nations such as China and the United States will compete for AI leadership, but most countries will access AI in a secondary manner. Given the low marginal cost of replication of AI models, the unknowable nature of future AI developments and the risks of misuse of these technologies, Bremmer and Suleyman (ibid.) argue for establishment of a three-part global AI governance regime. First, they suggest establishment of a global scientific body, similar to the UN Intergovernmental Panel on Climate Change, to advise governments on emerging AI challenges. Second, they propose an international body to manage tensions among major AI powers (for example, China and the United States), to establish guardrails and monitoring, and to prevent proliferation of dangerous advanced AI systems. Third, the authors propose a technocratic body for AI risk — a Geotechnology Stability Board21— to engage national regulators, international standards bodies and private sector multinational stakeholders in preventing or responding to AI crises, and governing open-source AI. Across these three bodies, regulatory mechanisms might be considered such as “know your customer” transparency standards, licensing requirements, safety testing protocols and product registration and approval processes, among others. According to Bremmer and Suleyman (ibid., 36) the AI governance regime should “identify and mitigate risks to global stability without choking off AI innovation and the opportunities that flow from it.” They call this “technoprudentialism,” an approach that would be guided by common principles: an appropriate degree of precaution (first, do no harm); regulatory agility in anticipation of technological changes; inclusivity (engaging AI stakeholders including governments, private sector AI firms and civil society experts such as scientists, ethicists, trade unions, among others); impermeability, meaning blanket coverage globally; 21 Citing the example of the Financial Stability Board, Robert Fay and Rohinton Medhora (2021) make a similar proposal for a Digital Stability Board that would operate with respect to regulation of the international digital economy more broadly. and targeting actions depending on the risks in specific contexts (rather than one-size-fits-all). Ho et al. (2023) call for an Advanced AI Governance Agency to set international standards, support implementation of standards and monitor compliance. They suggest establishment of institutional capacity for several further functionalities, either via stand-alone institutions or in some combination: an AI safety project to conduct or support research; a scientifically oriented commission on frontier AI to identify opportunities and risks, and contribute to policy consensus around these; and a public-private frontier AI collaborative to develop this technology, manage dual-use applications, educate, and support diffusion and access. In developing these functionalities care will be required to avoid encumbrances of the type that hinder the institution’s agility and responsiveness. Leading AI-Intensive Firms: A Few Stylized Facts A few comparative, stylized facts may help to provide a sense of the differentiating characteristics of AI-intensive, innovative firms. In the search for such firms, the study employs two separate firm-level analyses drawing on two independent enterprise data sets that cover a broad range of industries and geographies. These cover two distinct segments from the population of innovative enterprises. The first covers medium and large publicly listed, R&D-expenditure leading firms. The second covers unicorn firms — successful start-ups that have achieved market valuations of approximately US$1 billion or more. Each of these two pools of firms is screened to identify firms that are likely to be among the leaders in AI technology development. The screens are based on published assessments from independent industry experts and each firm’s membership in relevant industry bodies. The results are illustrative rather than strictly statistically representative. The analysis here focuses on firms that have already achieved scale in the AI field. This matters for some dimensions of AI innovation 10 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt including, for example, the costly development and training of original LLMs. While there is considerable potential for smaller firms to innovate and build AI application and serviceprovision businesses, scale may confer advantages in some fundamental areas of development (for example, foundation model development). Consideration of such a selection of large firms provides an opportunity to examine areas of concentration, which in turn may have implications for diffusion of some types of AI innovation. AI Leaders among the Top Firms for R&D Expenditure Globally The first assessment begins with consideration of the top 2,500 firms for R&D expenditure globally.22 The analysis is based on data from the “EU Industrial R&D Investment Scoreboard,” which is produced annually under the auspices of the European Commission (Grassano et al. 2022). In addition to R&D expenditure, this EU publication presents a range of performance indicators for each firm such as net sales, employment, capital expenditure and profitability. Each firm’s country of registry and industry sector are also noted.23 Across all fields, these firms invested collectively a total of US$1.3 trillion in R&D during 2021.24 This represents about 86 percent of the estimated total global business expenditure on R&D that year,25 and roughly half of the total global innovation expenditure from all sectors, including business, government, academia and non-profit institutions.26 22 R&D covers basic research, applied research and experimental development. 23 See Table 1 notes for more information on the industry classification employed by the EU study team. 24 See Nicola Grassano et al. (2022); an online overview is available at https://iri.jrc.ec.europa.eu/scoreboard/2022-eu-industrial-rd-investmentscoreboard#block-ecl-theme-page-title. 25 Ibid. 26 The World Bank provides data on R&D expenditure as a percentage of GDP and also provides GDP value (current US$). In 2020, this amounted to 2.63 percent of global GDP (latest year available, United Nations Educational, Scientific and Cultural Organization [UNESCO] data as reported by the World Bank). Global GDP was US$96.88 trillion in 2021 (World Bank data). R&D includes capital and current expenditures in the four main sectors: business enterprise, government, higher education and private non-profit. The value of global R&D from all sources totalled about US$2.55 trillion. This data is available from the World Bank at https://data.worldbank.org/indicator/GB.XPD.RSDV.GD.ZS?name_ desc=false (R&D as a percentage of GDP) and at https://data.worldbank.org/indicator/NY.GDP.MKTP.CD?name_ desc=false (GDP, value). In screening for leading AI-intensive firms among the top 2,500 R&D expenditure leaders, the analysis initially employs three screens: a listing of leading AI innovative firms from a widely read technology and business publication (Enterprise Weekly, eWeek, “Top 100+ Artificial Intelligence Companies 2023” listing, as of May 29, 202327); firms included in Schwab’s “AI Thematic Research” listing as of August 18, 202328; and publicly listed firms in HSBC’s Global Research equity coverage for AI firms in China.29 AI intensity in a firm is taken here to mean that the firm’s AI technology is expected to contribute significantly to the firm’s own economic performance, and that the firm’s AI innovation or application of AI innovation may have economically important implications for other firms in terms of their own performance (for example, whereby some may benefit due to inbound AI technology transfer or imitation, and others such as direct competitors may find themselves at a disadvantage). The three screens were selected based on two principal considerations: to tap into the firm-level expertise of a broad selection of professional industry analysts tracking firms in a fast-evolving sector to get a sense of their economic potential; and to ensure broad geographic coverage of leading AI firms across North America, Europe and Asia (including China). Partial confirmation of the screening selection results was conducted by cross checking against AI industry group membership lists for the USanchored Partnership on AI as of August 202330 and the China-anchored Artificial Intelligence Industry 27 The current eWeek AI company listing is available at www.eweek.com/ artificial-intelligence/ai-companies/. In the May 2023 edition employed for the present research, based on the expert opinion of the editorial team, a list of firms considered to be leading in each segment of their taxonomy of AI sector firms was assembled. The taxonomy was as follows: conversational AI, cybersecurity, education, enterprise majors, financial services, generative AI, giants, health-care AI, pioneers, retail and AI, robotic process automation, and visionaries. Per the eWeek taxonomy, “visionaries” are AI start-ups that are “closer to the edge, inventing the generative AI landscape in real time”; “pioneers” are forward-looking AI companies inventing and supporting the systems that are propelling AI development. 28 Schwab research data is available to clients at www.schwab.com/clienthome. 29 HSBC Global Research references include Frank He (2023) for AI software and services, and Charlene Liu et al. (2023) for AI infrastructure. 30 According to the “About us” page on its website, the Partnership on AI is a non-profit partnership of academic, civil society, industry and media organizations creating solutions so that AI advances positive outcomes for people and society. The membership list is available via the Partnership on AI at https://partnershiponai.org/work/. 11AI Innovation Concentration and the Governance Challenge Alliance as of March 2020.31 Both industry groupings have international membership. In six instances, firms from among the top 2,500 R&D innovators were selected for inclusion in the illustrative list of AI leaders based on their membership in one of these AI industry groups and their significant AI developments as discussed elsewhere in the trade press: Apple and Samsung due to their work on AI chips and consumer-facing AI innovation, as well as their scale; Synposis due to innovation in AI chip design software; Seagate due to AI storage innovation; China Telecom due to AI platform and LLM innovation; and Nokia for AI innovation in telecoms. (Further details of the screening are presented in Appendix 1, Table A.1, including a note on firms omitted due to data limitations in the EU listing of the top 2,500 R&D innovators.) AI Leaders: Findings and Caveats The screening process identifies 57 leading AI innovator firms, 56 of which are listed on major stock exchanges (see Table 1 and Appendix 1, Table A.132). One striking result is the geographic concentration of these firms (see Table2). Fully two-thirds of these firms are based in the United States. Among the other geographies, China accounted for less than one in seven of the listings and the European Union for less than one in 10. Collectively in 2021, these firms recorded R&D expenditure of US$264 billion, with some two-thirds of that expenditure by US-registered firms. Firms in China accounted for just under one-fifth of the total R&D expenditure and the EU for just one-sixteenth. Together, the 57 leading AI-intensive firms in the panel accounted for roughly one-tenth of global investment in R&D from all sources. While the data does not permit an assessment of the shares of panel firms’ R&D with an explicit focus on AI, it is nonetheless 31 The Artificial Intelligence Industry Alliance (AIIA) website is available at http://aiiaorg.cn/index.php?m=alliance&c=index&a=structure. The version of the membership list employed in the present study was developed by the Center for Security and Emerging Technology (Luong and Arnold 2021). The authors describe the alliance as follows: “Through the AIIA, the Chinese government aims to foster collaboration among local governments, academic institutions, and companies. In some cases, the Chinese state uses the AIIA to ‘pick winners,’ choosing among favored companies in the AI industry to receive government subsidies” (ibid., 1). 32 For detailed firm-level performance indicators for these firms see Grassano et al. (2022), and the underlying data set available from the EU website at https://iri.jrc.ec.europa.eu/scoreboard/2022-eu-industrialrd-investment-scoreboard#field_reportscoreboard. clear that this panel of leading AI-intensive firms delivers research at a globally significant scale. With respect to industry (see Table 2), the concentrations in the Industry Classification Benchmark (ICB) sectors “Technology Hardware & Equipment” and “Software & Computer Services” are in line with prior expectations. However, the ranking includes quite a few large and diversified corporations (for example, Alibaba, Alphabet, Apple, Baidu, Meta, Microsoft, Samsung and Sony) with notable product offerings beyond their listed industry, including other fields where AI applications might be employed (such as media or leisure goods). Alongside global multi-trillion USdollar behemoths such as Apple and Microsoft, a number of smaller firms with a particular focus on AI are included in the rankings (see Figure1). One in three had a market capitalization of less than US$25 billion (Grassano et al. 2022). Thus, while large firms dominate the leading AIintensive grouping (providing an upward shift to the averages), there is a significant subset of smaller firms that have gotten traction. Table 3 provides comparisons between the leading AI-intensive firms, the remainder of the software and computer services firms (excluding the leading AI-intensive firms33), and the top 2,500 R&D firm sample as a whole. Considering the median scores for each group for the year 2021, the assessment reveals that the median leading AI-intensive firm is substantially larger than the median firm among the top 2,500 in absolute terms with respect to net sales (3.8x), employment (3.1x), R&D expenditure (8.4x), capital expenditure (capex) (3.3x), and operating profits (6.0x).34 This is due in part to the presence of very large and diversified corporations in the AI rankings. Development of original AI models can be costly in terms of computing requirements and firm scale may provide an advantage in this regard. Larger firms may also have a broader range of in-house use cases to fuel AI demand. This is not to say that smaller AI intensive firms are excluded from the market. A number of smaller firms are proving competitive in adaptation of existing LLMs for particular uses and in developing AI-driven applications. In some 33 As noted in Table 2, most, but not all, of the leading AI-intensive firms are in the ICB section “Software & Computer Services.” 34 In this notation, “x” signifies “times.” For example, net sales of the median leading AI-intensive firm are 3.8 times those of the median top 2,500 R&D firm. 12 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt cases, these smaller firms may compensate for lack of scale by drawing upon open-source inputs. A further confirmation of the importance of scale to some segments of the AI sector can be found in Table 3 by considering the gap between the median leading AI-intensive firm and the median software and computer services firm (excluding leading AI-intensives). This gap is even larger than for the top 2,500 R&D firm sample as a whole. The median software and computer services firm (excluding the AI intensives) has lower net sales, lower employment, less capital expenditure and much lower operating profits than the median for the top 2,500 as a whole. Only with respect to R&D expenditure does the median software and computer service firm (excluding AI intensives) outperform the median firm among the top 2,500 sample as a whole. Indeed, the median R&D intensity (R&D expenditure relative to net sales) for both the leading AI-intensive firms and the remainder of the “Software & Computer Services” sector is much greater than for the top 2,500 sample as a whole (respectively 2.6x and 3.0x). Caveats The data employed in this assessment may be subject to certain biases. There are two important caveats: First, the EU team depends on the public availability of R&D expenditure data. Only firms that reveal this information are covered. This results in coverage primarily of publicly listed companies, though even among publicly listed firms the disclosures vary. Coverage is better for firms in the advanced economies, which tend to have better corporate disclosure of information on R&D. Thus, some potentially important AI innovators may be left off the European Union’s ranking. According to the chosen AI-screening criteria, among large developers and innovative users of AI, some significant omissions include Amazon, Capital One, GE HealthCare, JD.com, McDonald’s, Lowe’s and Wipro. Second, some leading AI innovators are actually small in terms of the scale of R&D expenditure and thus may not meet the EU top 2,500 criteria, even if they are publicly listed and disclose R&D expenditure information. Among smaller publicly listed firms identified in the screens as leading in terms of AI technology but missing from the EU rankings one finds: Darktrace PLC, Innovative Eyewear Inc., Kore Group Holdings, Nano Dimension LTD, SoundHound AI Inc., and Xometry Inc. As a consequence of these caveats, this firm-level assessment may be considered as illustrative but not necessarily fully representative. AI Leaders among the Unicorns There is an active insurgency in the AI-intensive firm segment. A significant portion of the innovation in AI has been fuelled by start-ups challenging — Table 1: Leading AI-Intensive Firms among Global R&D Expenditure Leaders, Publicly Listed, 2021 Company Country Industry (ICB Sector) Company Country Industry (ICB Sector) Accenture Ireland Support Services Intuit United States Software & Computer Services Adobe United States Software & Computer Services Medtronic Public Limited Ireland Health Care Equipment & Services Alibaba Group Holding China Software & Computer Services Meta United States Software & Computer Services Alphabet United States Software & Computer Services Microsoft United States Software & Computer Services Altair Engineering United States Software & Computer Services NetEase China Software & Computer Services Alteryx United States Software & Computer Services NICE Israel Software & Computer Services Ambarella United States Technology Hardware & Equipment Nokia Finland Technology Hardware & Equipment Apple United States Technology Hardware & Equipment Nvidia United States Technology Hardware & Equipment 13AI Innovation Concentration and the Governance Challenge Company Country Industry (ICB Sector) Company Country Industry (ICB Sector) Baidu China Software & Computer Services Oracle United States Software & Computer Services Broadcom United States Technology Hardware & Equipment Palo Alto Networks United States Software & Computer Services Butterfly Network United States Health Care Equipment & Services Pegasystems United States Software & Computer Services C3.ai United States Software & Computer Services Rockwell Automation United States Industrial Engineering Cambricon Technologies China Technology Hardware & Equipment Salesforce United States Software & Computer Services Cerence United States Software & Computer Services Samsung Electronics South Korea Electronic & Electrical Equipment Ceva United States Mobile Telecommunications SAP Germany Software & Computer Services Check Point Software Technologies Israel Software & Computer Services Seagate Technology Ireland Technology Hardware & Equipment China Telecom China Technology Hardware & Equipment ServiceNow United States Software & Computer Services Coherent United States Electronic & Electrical Equipment Shutterstock United States Media CrowdStrike Holdings United States Software & Computer Services Snowflake United States Software & Computer Services Dell Technologies United States Technology Hardware & Equipment Sony Japan Leisure Goods Duolingo United States Software & Computer Services SS&C Technologies United States Software & Computer Services Fortinet United States Software & Computer Services Stryker United States Health Care Equipment & Services Hewlett Packard Enterprise United States Software & Computer Services Synaptics United States Technology Hardware & Equipment Huawei Investment & Holding China Technology Hardware & Equipment Synopsys United States Software & Computer Services IBM United States Software & Computer Services Tencent China Software & Computer Services iFlytek China Software & Computer Services Thomson Reuters Canada Support Services Informatica United States Software & Computer Services UiPath United States Software & Computer Services Infosys India Software & Computer Services Zscaler United States Software & Computer Services Intel United States Technology Hardware & Equipment Source: Grassano et al. (2022); author’s tabulations. The underlying data set is available from the EU website at https://iri.jrc.ec.europa.eu/scoreboard/2022-eu-industrial-rd-investment-scoreboard#field_reportscoreboard. Notes: Huawei Investment & Holding is employee-owned. Also, the ICB taxonomy is employed here. The firm FTSE Russell manages the taxonomy. For details, see www.lseg.com/en/ftse-russell/industry-classification-benchmark-icb and Grassano et al. (2022, 11). Table 1 (continued) 14 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt Table 2: Counts of Top Publicly Listed AI Innovator Firms, by Country of Registry and Sector, 2021 Country Industry (ICB Sector) Canada 1 Electronic & Electrical Equipment 2 China 8 Health Care Equipment & Services 3 Finland 1 Industrial Engineering 1 Germany 1 Leisure Goods 1 India 1 Media 1 Ireland 3 Mobile Telecommunications 1 Israel 2 Software & Computer Services 34 Japan 1 Support Services 2 South Korea 1 Technology Hardware & Equipment 12 United States 38 Total AI innovator firms 57 Total AI innovator firms 57 Of which registered in EU member countries 5 Source and Notes: Refer to Table 1 source and notes. and then in some cases eventually collaborating with— the larger incumbents. Among the most well-known, for example, are OpenAI (developer of the ChatGPT model, now affiliated with Microsoft) and ByteDance (developer of TikTok and a government-approved LLM called Yunque; ByteDance is a private firm with a public listing pending). The most successful AI-intensive start-ups, such as these two, can grow to become unicorns many times over (be valued at many US$ billions). In this second firm-level assessment, the analysis shifts to consider the global database of 1,220 unicorns developed by CB Insights, a list covering firms from a broad range of sectors. By crossreferencing this database against the roster of AI leaders published by eWeek (Maguire 2023), the analysis develops a working list of top AI-intensive unicorns (see Table 4). Based on references in the literature, the analysis added two further start-ups. The first is Epic Games, a US-based unicorn affiliated with Tencent. This firm maintains innovative transparency provisions for content management. Generative AI content is welcomed from vendors on the firm’s Unreal Engine Marketplace, but must be labelled as such. Vendors may also exclude their content from use by others in AI training (Epic Games 2023). The second additional start-up is Zhipu AI, an innovative start-up selling access to its AI models as a service. Zhipu AI also has a LLM that received government approval in China in August 2023 (Liu et al. 2023). Unicorns: Findings and Caveats One striking characteristic of this list of 22 unicorns is the geographic concentration. All but four of these unicorns are US-based firms, the others being based in China (two) and the United Kingdom (two). Looking a bit closer, the extent of the geographic concentration is revealed. Only a few cities within these three countries actually host these unicorns (see Figure 2). This view found 10 unicorns emerging in the San Francisco-Mountain View-San Jose corridor of California. Another five are based in New York City, with Boston, Massachusetts, Cary, North Carolina, and Austin, Texas, accounting for the remaining US entries on the list. London accounts for two and Beijing for two. In terms of sector, nearly two-thirds (14) of the unicorns are focused on enterprise technology. Three are concerned with consumer and retail applications, three with media and entertainment, and two with health care and life sciences. Most of the current AI leaders among the unicorns are focused on products for businesses (business-to-business [B2B]), even though some of the B2B products will be employed in consumer-facing applications. A few are delivering direct-to-consumer products such as TikTok (ByteDance), Fortnite (Epic Games) or ChatGPT and DALL-E (OpenAI). 15AI Innovation Concentration and the Governance Challenge Counts Median Scores Number of firms Number of countries Number of regions Number of industry (ICB) sector names Median world rank R&D 2021 (US$M) R&D one-year growth (%) Net sales (US$M) Net sales one-year growth (%) R&D intensity (%) AI-intensive firms 57 10 5 9 200 1,168.1 18.3 10,061.7 18.0 16.9 Software and computer services firms (excluding AIintensive) 302 23 5 1 1143 156.7 22.7 782.5 21.8 19.3 Top 2,500 R&D firms (full sample) 2,500 41 5 38 1251 139.4 15.7 2,627.2 16.9 6.4 Median Scores Capex (US$M) Capex one-year growth (%) Capex intensity (%) Operating profits (US$M) Operating profits one-year growth (%) Profitability (%) Employees Employees, one-year growth (%) Market capitalization (US$M) Market capitalization one-year growth (%) AI-intensive firms 341.4 5.6 3.1 1,297.6 26.8 14.0 23,437 11.3 53,759.0 36.7 Software and computer services firms (excluding AIintensive) 18.2 11.3 2.0 4.6 12.0 1.2 3,610 15.0 5,682.4 17.2 Top 2,500 R&D firms (full sample) 104.3 13.2 4.1 216.1 21.1 8.0 7,522 5.3 5,200.4 26.6 Source: Grassano et al. (2022); underlying data set is available from the EU website at https://iri.jrc.ec.europa.eu/scoreboard/2022-eu-industrialrd-investment-scoreboard#field_reportscoreboard; author’s tabulations. US$ figures are converted from euros using the US Federal Reserve annual exchange rate (G.5A), available at www.federalreserve.gov/releases/g5a/current/. Note: As defined in Grassano et al. (2022), the regional taxonomy includes China, the European Union, Japan, the United States, and the rest of the world. Table 3: Counts and Median Scores for the Group of Leading AI-Intensive Firms versus Other Groupings, 2021 16 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt Figure 1: Top Publicly Listed AI-Intensive Firms with Respect to R&D Expenditure, Ranked by Market Capitalization 0 500,000 World rank for R&D expenditure Market capitalization (US$M, 2021) 1,000,000 2,000,000 2,500,000 3,000,0001,500,000 Apple Microsoft Meta Alphabet Tencent NVIDIA Alibaba Group Holding Samsumg Electronics Adobe Salesforce Oracle Accenture Intel Broadcom Medtronic Public Limited SAP Intuit ServiceNow IBM Sony Stryker Infosys Snowflake NetEase CrowdStrike Holdings Thomson Reuters Fortinet Synopsys Palo Alto Networks Baidu Zscaler Rockwell Automation Nokia Dell Technologies UiPath Seagate Technology Hewlett Packard Enterprise SS&C Technologies NICE Check Point Software Technologies iFlytek Pegasystems Synaptics Coherent C3.ai China Telecom Alteryx Shutterstock Cerence Ambarella Altair Engineering Butterfly Network Ceva Source: Grassano et al. (2022); underlying data set is available from the EU website at https://iri.jrc.ec.europa.eu/ scoreboard/2022-eu-industrial-rd-investment-scoreboard#field_reportscoreboard; author’s tabulations. Note: The EU database does not include market capitalization data for Cambricon Technologies, Duolingo, Huawei Investment & Holdings or Informatica. 23AI Innovation Concentration and the Governance Challenge Table 5, Part 2: Counts of Policy Initiatives as of 2021 in Relation to AI in an Illustrative Group of Other Countries Sample Countries Non-sample Countries Policy Instrument Type Median Brazil Indonesia Singapore South Africa Vietnam AI coordination and/or monitoring bodies 2.0 1 0 0 0 0 AI computing and research infrastructure 2.5 0 0 0 0 1 AI skills and education 1.0 1 0 0 0 0 AI use in the public sector 1.0 13 0 6 0 0 Centres of excellence grants 0.5 1 0 0 0 0 Data access and sharing 1.0 0 0 1 0 0 Emerging AI-related regulation 5.0 2 0 6 0 1 Equity financing 0.0 0 0 0 0 0 Fellowships and postgraduate loans and scholarships 0.0 0 0 0 0 0 Grants for business R&D and innovation 0.5 0 0 1 0 0 Indirect financial support 0.0 0 0 0 0 0 Institutional funding for public research 0.0 1 0 1 0 0 Knowledge transfers and business advisory services 1.0 1 0 2 0 0 Labour market policies 0.0 0 0 0 0 0 Labour mobility regulation and incentives 0.0 0 0 0 0 0 National strategies, agendas and plans 7.0 5 1 9 0 2 Networking and collaborative platforms 4.0 0 0 2 1 0 Procurement programs for AI R&D and innovation 1.0 0 0 0 0 1 Project grants for public research 2.5 1 0 1 0 1 Public awareness campaigns and civic participation activities 1.0 0 0 2 0 1 Public consultations of stakeholders or experts 5.0 2 1 1 1 1 Regulatory oversight and ethical advice bodies 1.5 0 0 1 1 0 Standards and certification for technology development and adoption 0.0 1 0 4 0 0 Science and innovation challenges, prizes and awards 0.0 0 0 0 0 0 Total count 43.5 29 2 37 3 8 Memo Items Policy initiatives in the OECD.AI policy database that mention “innovation” in the objective description 9.0 19 2 11 1 2 Share of initiatives tagged for having a mention of “innovation” in the objective description (percentage) 20.4 65.5 100.0 29.7 33.3 25.0 Policy initiatives in the OECD.AI policy database addressing transparency issues (e.g., in policy formation and regulatory processes or in the operation of algorithms) 11.5 11 2 10 0 0 Share of initiatives tagged in relation to transparency (percentage) 32.4 37.9 100.0 27.0 0.0 0.0 Sources and Notes: See Table 5, Part 1, sources and notes. 24 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt impediments to diffusion of AI innovation may also be addressed as part of these efforts.40 DPA: Machine Learning and AI Policy Activity In order to get a better sense of the recent policy dynamics around the acceleration in AI, the analysis turns to a more current resource for tracking AI policy activity. This is DPA, an initiative of the St. Gallen Endowment for Prosperity through Trade. It covers policy developments in G20 countries, EU member states and Switzerland. The sample employed for the present analysis is for actions relevant to machine learning41 and AI. The actions tracked include laws, orders, standards and guidelines regulating the development and use of AI systems, as well as related regulatory actions and policy development work. Activities may focus on one or more policies and a single policy may be the object of one or more activities over time. The database is updated in an ongoing manner with the support of customized software. The structure of the DPA database is conducive to consideration of policy activity over time for recent years starting in 2020. Table 6 provides an illustration of the acceleration in policy activity between two periods, 2020–2021 and 2022– November 2023. From the initial period, all of the sample countries hosting leading AI-intensive firms were already engaged in machine learning and AI policy activities. The United States led the group with 15 incidents of policy activity, followed by the European Union (10), the United Kingdom (eight) and China (seven). A few international country groupings also took some steps in the first two years covered here (2020–2021). In the second period, from 2022 to November 2023, the volume of activity accelerated by 4.2 times overall. The leading AI-intensive firm host countries all stepped up policy activity, including the United States (to 99 activities), China (to 26), the European Union (to 20) and the United Kingdom (to 20). The number of actions across international borders also ramped up, from five in the initial two years to 37 in the period from 2022 to November 2023. 40 For more on this perspective of regulatory reform, see the OECD efficient regulation principles for market openness (Tsai et al., 2011, 18, box 1) and the detailed discussion of the application of these principles in the case of Israel (ibid., 18–39). 41 Machine learning involves use of a model trained and refined on large data sets to classify information, make predictions and generate content. It enables computers to learn without being explicitly programmed. Each of these international actions engaged one or more countries from the sample. The subject matter covered increased in scope and intensity (see Table 7). During 2020–2021, 12 subject areas were tackled. In 2022–2023, this rose to 30. In both periods, “Algorithm design and technical standards” exhibited the most activity. “Data protection regulation” activity rose dramatically from the first period to the second. Also during the second period, “Cyber security regulation,” “Content moderation regulation,” “Intellectual property (copyright and patents)” and “Consumer protection regulation” all rose to fill out the top ranks of policy areas. Moreover, the increased volume of policy activity led to substantive results. Activities for policies adopted rose from 40 during the first biennium, to 114 during the second. A further 35 actions concerned policies in force during 2022–2023. In addition, the pipeline of policy activity points to a continued high volume of activity going forward. During 2022–2023, some 45 instruments were at some point in the consultation process and 80 were under deliberation. Among the leading issues for activities concerning policy instruments under deliberation, “Algorithm design and technical standards” remains in first place and “Data protection regulation” places a strong second. Findings from Analysis of DPA Tracking The DPA tracking highlights the responsiveness of policy makers to the acceleration in AI innovation, with the frequency of policy initiatives increasing as well as the country coverage. Some of the subject areas frequently targeted, such as technical standards and IP rights, can have an influence on diffusion of technology across borders. Alignment in regulation in these areas can facilitate trade in products and services by permitting developers to aim for a more uniform product with less tailoring required, potentially also reducing the burden of testing and certification where required. The increased coverage of protection of IP rights and data protection may also enable developers to sell AI products into an expanded range of markets (for example, if the reforms enable developers to better protect their products from abuses or if the reforms reduce uncertainty with respect to liability). Also, concerning regulatory alignment, it is encouraging to see the growing number of country groupings taking joint policy initiatives, often including the major economies. 25AI Innovation Concentration and the Governance Challenge Table 6: Dynamics of Recent Machine Learning and AI Policy Developments, January 2020– November 2023 (Counts) 2020–2021 2022–2023 Country or Group of Countries Policy Activity Country or Group of Countries Policy Activity — Argentina 5 Australia 3 Australia 6 — Austria 1 Brazil 2 Brazil 7 Canada 2 Canada 9 China 7 China 26 — Denmark 1 European Union 10 European Union 20 — France 6 Germany 3 Germany 9 — Greece 1 Hong Kong 1 — — Hungary 1 India 2 India 3 Indonesia 1 Indonesia 1 Italy 1 Italy 4 — Japan 8 — Netherlands 3 — New Zealand 2 — Peru 2 — Poland 1 Russia 1 Russia 1 —Saudi Arabia 1 — Singapore 4 —South Africa 1 South Korea 4 South Korea 8 Spain 1 Spain 5 Switzerland 1 Switzerland 2 Taiwan 1 — Türkiye 2 — — Ukraine 1 —United Arab Emirates 1 United Kingdom 8 United Kingdom 20 United States 15 United States 99 Australia, European Union, India, Japan, Mexico, New Zealand, Singapore, South Korea, United Kingdom, United States 1 — Canada, United Kingdom, United States 1 — Chile, New Zealand, Singapore, South Korea 1 — UNESCO membership 2 — 26 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt 2020–2021 2022–2023 Country or Group of Countries Policy Activity Country or Group of Countries Policy Activity — Albania, Argentina, Brazil, Chile, China, Colombia, Ecuador, Egypt, Estonia, Ethiopia, Finland, France, Germany, Guyana, India, Israel, Italy, Japan, Kenya, Mexico, Netherlands, New Zealand, Pakistan, Peru, Portugal, Russia, Saudi Arabia, Senegal, Singapore, South Africa, South Korea, Spain, Suriname, Switzerland, Trinidad & Tobago, United Arab Emirates, United Kingdom, United States, Zimbabwe 1 —Ibero-American Network: Andorra, Argentina, Brazil, Cape Verde, Chile, Colombia, Costa Rica, Dominican Republic, Ecuador, El Salvador, Guatemala, Honduras, Mexico, Nicaragua, Panama, Paraguay, Peru, Portugal, Sao Tome & Principe, Spain, Uruguay 1 — Argentina, Australia, Canada, Colombia, Hong Kong, Mexico, Morocco, New Zealand, Norway, Switzerland, United Kingdom 1 — Australia, Austria, Belgium, Brazil, Bulgaria, Canada, Chile, China, Croatia, Cyprus, Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, India, Indonesia, Ireland, Israel, Italy, Japan, Kenya, Latvia, Lithuania, Luxembourg, Malta, Netherlands, Nigeria, Philippines, Poland, Portugal, Romania, Rwanda, Saudi Arabia, Singapore, Slovakia, Slovenia, South Korea, Spain, Sweden, Switzerland, Türkiye, Ukraine, United Arab Emirates, United Kingdom, United States 1 — OECD: Australia, Austria, Belgium, Canada, Chile, Colombia, Costa Rica, Czechia, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Israel, Italy, Japan, Latvia, Lithuania, Luxembourg, Mexico, Netherlands, New Zealand, Norway, Poland, Portugal, Slovakia, Slovenia, South Korea, Spain, Sweden, Switzerland, Türkiye, United Kingdom, United States 2 — Australia, Canada, Germany, Netherlands, New Zealand, United Kingdom, United States 1 — Australia, United Kingdom 1 — G7: Canada, France, Germany, Italy, Japan, United Kingdom, United States 4 —Canada, Japan 2 —China, Saudi Arabia 1 — European Union, India 1 — European Union, South Korea 1 — European Union, United States 3 —Germany, Japan 1 — India, Israel 1 — India, United Kingdom 1 — India, United States 1 —Indonesia, Japan 1 — Japan, United Kingdom 2 — Germany, South Korea 1 — Singapore, South Korea 1 — South Korea, United States 1 Table 6 (continued) 27AI Innovation Concentration and the Governance Challenge Table 6 (continued) Table 7: Nature of Recent Machine Learning and AI Policy Developments, January 2020– November 2023 (Counts) Panel A: Policy Instruments 2020–2021 2022–2023 Policy Instruments Policy Activity Policy Instruments Policy Activity Algorithm design and technical standards 23 Algorithm design and technical standards 99 — Authorization of goods or services 8 — Border measure, n.e.s. 1 — Business registration requirement 10 — Competition authority governance 1 — Consumer protection regulation 11 — Content moderation regulation 12 — Content remuneration regulation 1 Content-related regulation and IP, n.e.s. 1Content-related regulation and IP, n.e.s. 4 — Cybersecurity regulation 14 — Data protection authority governance 7 Data protection regulation 9 Data protection regulation 58 Direct taxes, including digital service taxes 1 — Export licensing requirements 2 Export licensing requirements 5 Financial grant 5 Financial grant 6 Instrument unspecified 13 Instrument unspecified 5 — Intellectual property: copyright 6 Intellectual property: patents 4 Intellectual property: patents 6 Merger control regulation 1 Merger control regulation 1 MNE: Financial incentive 1 — Operating condition, n.e.s. 9 Operating condition, n.e.s. 4 — Operational licence requirement 1 — Production subsidy 1 — Public procurement access 4 2020–2021 2022–2023 Country or Group of Countries Policy Activity Country or Group of Countries Policy Activity — Singapore, United Kingdom 1 — Singapore, United States 2 — Ukraine, United Kingdom 1 — United Kingdom, United States 3 Total 70 Total 296 Sources: DPA (2023); author’s tabulations using the underlying data. Notes: “—” indicates that there is no comparable activity entry for the cell during the time period covered by the column. Also note that the DPA Activity Tracker provides information on recent developments in legislatures, judiciaries and the executive branches of the G20, EU member states and Switzerland. Among the AI leader countries most are covered. The DPA tracking does not have separate listings for Ireland or Finland, but they are partly covered via the EU listings. Israel is only covered for its international accords, not for its domestic policy activity. 28 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt Panel A: Policy Instruments 2020–2021 2022–2023 Policy Instruments Policy Activity Policy Instruments Policy Activity — Public procurement regulation, n.e.s. 1 — Quality of service requirements 7 — Registration and licensing, n.e.s. 1 Service access restrictions 1 Service access restrictions 1 — State aid, n.e.s. 7 — Unilateral conduct regulation 3 — Worker monitoring regulation 9 — Worker status classification regulation 2 Total 70 Total 296 Panel B: Status of Policy 2020–2021 2022–2023 Status Types Policy Activity Status Types Policy Activity Adopted 40 Adopted 114 Concluded 1 Concluded 3 In consultation 1 In consultation 7 In force 11 In force 35 — In grace period 1 Processing consultation 6 Processing consultation 38 Rejected 1 Rejected 12 Under deliberation 10 Under deliberation 80 — Under investigation 6 Total 70 Total 296 Sources and Notes: See Table 6. In addition, MNE = multinational enterprise; n.e.s. = not elsewhere specified. Table 7 (continued) Leading AI-Enterprise Host Country Regulatory Specifics As the global governance framework for AI is still under development, steps taken by the major economies (China, the European Union and the United States; see Appendix 2) may have a significant influence. This influence derives not only from the market power of policy and regulatory actions taken in their own domestic markets, but also from their influence in international bodies (for example, the OECD, the Asia-Pacific Economic Cooperation regional forum (APEC),42 the G7 and the G20),43 through their industry associations (for example, the AI Industry Alliance in China or the Partnership on AI and the Data and Trust 42 For information about APEC, see www.apec.org/About-Us/About-APEC/. 43 At the international level, for example, these nations were engaged in the UNESCO “Recommendations on the Ethics of AI in 2021” (UNESCO 2022) and the Bletchley Declaration on AI Safety agreed by most G20 countries and others in 2023 (GOV.UK 2023). Furthermore, in a newly established EU-US tech partnership (the Trade and Technology Council), the European Union and the United States are seeking to develop a mutual understanding on the principles underlining trustworthy and responsible AI (Madiega 2023). They both have endorsed the OECD’s Recommendation of the Council on AI, which lays out basic principles (OECD 2019). 29AI Innovation Concentration and the Governance Challenge Alliance in the United States), and through their trade measures (for example, US export controls affecting third parties). China has been a first mover in AI regulation. Canadian,44 EU and US regulators also have significant elements in place, often drawing on pre-existing law and regulations applicable to the digital economy more generally (Aaronson 2024). But, they are also moving forward vigorously with AI-specific measures. Given the potential for cross-border diffusion of AI technology, it is clear that inclusive international coordination is needed to address concerns (Bremmer and Suleyman 2023). This would apply, for example, with respect to matters ranging from traceability and safety to assignment of IPRs and protections (for example, for algorithms). Some aspects will be sector specific (for example, patient safety and privacy in health care), whereas others will be crosscutting (for example, provisions for regulators’ access and oversight of algorithms). At the same time, in acting to address such concerns, the international community may reduce impediments to responsible diffusion of AI technology. Such a result could promote more inclusive economic gains from this technology internationally (DSIT 2023). China’s first-mover status in AI regulation has given the authorities valuable hands-on experience to inform future regulation. It appears a range of motivations are involved: a goal of ensuring adequate controls on information content in line with government policy for political and social stability; a desire for transparency and adequate protections; and an objective of having a policy environment conducive to Chinese leadership in AI technologically and commercially, including with respect to ethical norms and standards (Sheehan 2023). China’s three-part regulatory framework (Cyberspace Administration of China [CAC] 2021, 2022, 2023) covers algorithmic recommendations, deep synthesis internet information services (for example, virtual reality, deep learning and generative sequencing algorithms) and generative AI services. The framework provides for registration of LLMs, with the first batch of 12 models reportedly being authorized in August 2023 (Liu et al. 2023). The EU authorities have sought to balance support for AI sector development with concerns about AI safety and conditions for a competitive market. 44 The case of Canada is covered in Appendix 2. A substantial landmark piece of legislation, the Artificial Intelligence Act, is in advanced stages of development and entry into force is anticipated by mid-year 2024 (for example, see Bamberg et al. 2024; Dentons 2023; Heikkilä 2023).45 This legislation includes a tiered AI risk management framework that would ban unacceptable risks (for example, models that employ harmful manipulative subliminal techniques), strictly regulate high-risk AI systems (for example, biometric identification and categorization of natural persons), impose transparency and operational obligations on limited risk AI systems (for example, synthetic output from generative AI systems will need to be labelled as such), and permit free use of low and minimal risk AI systems (for example, spam filters or video games), although potentially subjecting them to a future code of conduct. General purpose AI models (foundation models, LLMs) will entail additional compliance requirements to maintain and provide current technical documentation, system transparency, copyright compliance and information on training data. The AI Act also establishes an EU AI office in the European Commission to supervise and enforce certain aspects of compliance in cooperation with a new EU AI board of member country representatives, a new scientific advisory panel of independent experts and a new network of national competent authorities in each member state. Penalties for firm-level violations of compliance requirements could be substantial, ranging from 1.5 percent to seven percent of annual turnover (subject to certain minimum and maximum thresholds). The AI Act will fit into a system of existing measures governing the digital economy in the European Union including the General Data Protection Regulation, the Cybersecurity Act, the Digital Services Act and the Digital Markets Act, among others. Likewise in the United States, the government has relied on existing legal authorities to discipline AI stakeholders. As US Federal Trade Commission Chair Lina M. Khan (2023, 1) noted, “There is no AI exemption to the laws on the books.” So, a crime such as fraud may remain a crime even if the fraud takes place using a new type of technology. With respect to AI-specific regulation, the United States had initially taken a lenient approach before advancing over the past two years in a stepwise fashion, allowing innovation in the private sector to thrive. It set out basic principles 45 See also European Parliament (2024). 30 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt for AI governance in October 2022, consulted with stakeholders during the first half of 2023, agreed a voluntary regime with 15 leading firms during summer 2023, then followed with a comprehensive Executive Order drawing on existing legislative authorities in October 2023. This measure imposes notification requirements on developers of dual-use foundation models.46 It also mandates federal agencies to proceed with responsible and strategic exploitation of AI systems, as well as administrative steps to support AI standards, tools and tests, human capital development, research and international cooperation, among other elements. Other economies engaged in leading-edge work on AI face the challenge of establishing their own domestic AI governance frameworks and defending national interests, while also striving for meaningful cooperation with relevant international institutions and the three major economies (China, the European Union and the United States). Canada provides a useful example (see Appendix 2). Domestically, it is working to complete legislation providing a basis for a principled, risk-based framework for regulation of AI. The Artificial Intelligence and Data Act (AIDA)47 would provide for further development of AI governance in the nation, while allowing flexibility to maintain coherence with key partner economies internationally, as feasible. Internationally, Canada is relatively well positioned with respect to some relevant digital economy matters covered via provisions of its free trade agreements with the European Union (the Comprehensive Economic and Trade Agreement, and related accords including the General Data Protection Regulation [GDPR] adequacy), with the United States and Mexico (CUSMA) and 10 Pacific Basin partners (CPTPP). For example, these accords can be supportive of AI development with respect to access to data. Canada is also supporting collaboration related to the digital economy (including AI) with 46 The term “dual-use foundation model” means an AI model that is trained on broad data; generally uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts; and that exhibits, or could be easily modified to exhibit, high levels of performance at tasks that pose a serious risk to security, national economic security, national public health or safety, or any combination of those matters (Biden 2023). Elliot Jones (2023) provides further elaboration on this type of model. 47 The House of Commons concluded its second reading of AIDA on April 24, 2023. As of January 31, 2024, it is currently under consideration before the Standing Committee on Industry and Technology. See LEGISinfo, “C-27, 44th Parliament,” available at www.parl.ca/legisinfo/en/bill/44-1/C-27. members of the APEC forum, the WTO and the OECD, and other international institutions. Conclusions AI Innovation Fuels Pressure on Regulators Over the past decade, the pace of technological advance in the AI sector has accelerated with respect to software architecture and algorithms, training data assembly and mobilization of powerful computing hardware, among other areas. The emergence of readily accessible generative AI has opened the door to broad use of the technology by the wider public. LLMs, in particular, have marked some substantial and highly visible progress. Over the past two years, the release of text-driven LLMs such as ChatGPT and Bard surprised some observers with the levels of functionality provided. While such AI advances are real enough, there is also an element of euphoria among some users, perhaps already anticipating the next steps in AI development (Murgia and Thornhill 2023). Indeed, despite AI having its early roots dating back to the 1950s, and despite the recent wave of progress in AI development, it is still early days in the AI revolution.48 Integration of AI into many business processes is only now scaling up. For example, increased application of AI in industrial automation and multitasking robotics has the potential to unlock significant additional productivity gains. There remain substantial unfulfilled AI objectives on the horizon. For example, the long-time AI community goal of delivering artificial general intelligence — that is, the human-equivalent capacity for perception, reasoning, inference and action in an AI system — remains to be achieved. With all that has been accomplished in the sector and the prospect of more powerful systems under development, regulators are indeed feeling some urgency in the need for improved AI governance. 48 For example, although use of embedded AI is widespread in US businesses, the US Census Bureau recently found that as of end of October 2023 only 3.8 percent of US businesses reported using AI to actually produce goods and services, though there is wide variation across sectors (Breaux and Dinlersoz 2023). 31AI Innovation Concentration and the Governance Challenge Concentration and Diffusion of AI Innovation At the outset of this paper, a central concern was highlighted to examine the concentration of AI innovation and to consider factors in the regulatory environment that may support or hinder its responsible diffusion. The illustrative assessment of leading AI-intensive firms (see the section “Leading AI-Intensive Firms: A Few Stylized Facts” above) confirmed the extent of concentration for at least two significant populations of AI-intensive firms: those AI-intensive firms in the upper echelon of global businesses for R&D expenditure (selected from among the top 2,500 firms globally for R&D expenditure in all fields); and those AI-intensive firms with success as a start-up in growing their valuation beyond US$1 billion (unicorns). In both cases, the assessment found a fairly high concentration geographically in AI innovators. The firms in these two populations represent just 11 nations, three of which are members of the European Union. The United States is home to a majority of the AI firms in these samples. The concentration of firms in just 11 host countries may in part be the result of agglomeration effects, whereby development in crowded technology centres may offer some advantages for advancement of AI development and diffusion. For example, thick labour markets may develop and offer large pools of sector-relevant talent. And, improved communication around innovation could emerge due to the proximity of stakeholders, thereby conferring further information advantages.49 Internationally, the availability of AI innovation depends in part on openness in channels for technology transfer. This may take place via such means as the sale or licensing of products and services, foreign direct investment, joint ventures and external staff training initiatives, among other possibilities.50 There are also opportunities for remote access to AI innovation through online hubs and software-as-a-service providers. From the literature review (see the section “Literature: Rapid Technological Development, Unsettled Scholarship” above), some factors associated with diffusion of technology can be identified. For 49 For example, see the discussion on agglomeration in Paul Krugman (1995). 50 For example, Park and Lippoldt (2014) provide an examination of these technology transfer issues from an IP perspective. example, availability of adequate protection of IP protection may play a role in facilitating transfers (for example, enabling identification of rights and obligations with respect to transferred intellectual assets). Establishment of an appropriate, aligned international regulatory regime may also help firms to manage risks arising from technology transfer by clarifying responsibilities with respect to safety, consumer protection, security, and promote reliability and trustworthiness in AI, among other issues. Indeed, regulatory misalignment entails costs that may inhibit technology transfer (see Box 2). In the absence of tighter international coordination, there is a risk that inconsistencies may emerge in the regulatory regime internationally and potentially even domestically (for example, between jurisdictions or ministries). To date, much of the AI regulatory activity has been at the national level as highlighted in the regulatory review section above. The review points to actions under way in a variety of areas potentially relevant to diffusion of AI innovation. For example, already as of 2021, seven of the reviewed economies — five sample economies (Canada, China, the European Union, Ireland and the United States) plus two non-sample countries (Brazil and Singapore) — had policy initiatives concerning standards and certification for technology development and adoption. While international cooperation has supported some peer learning and some convergence (for example, among OECD members), some divergences have appeared. For example, content moderation approaches in the United States or the European Union are inconsistent with the controls on content in China with respect to political and social stability. Among the first 12 LLMs registered under Chinese regulatory requirements in 2023, all originated with Chinese firms and institutions. On the other hand, certain of the non-sample countries considered in the review such as Indonesia, South Africa or Vietnam appear to be lagging in their AI regulatory development in terms of the number of actions under way. Such gaps in regulatory regimes could increase uncertainty around the future conditions for AI businesses and potentially have an inhibiting effect on technology transfer. The national provisions have been complemented in some cases by AI-relevant international accords that may help to ensure some convergence in regimes. For example, this can be the case in digital economy provisions in some trade agreements. 32 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt These can affect matters such as limiting recourse to data localization, prohibiting imposition of customs duties on data transmission and digital products, ensuring free cross-border transfer of personal data, requiring protection for consumers’ personal data and requiring cybersecurity measures (for instance, this is the case with the CPTPP and CUSMA) (Suominen 2021). The EU AI Act goes further to standardize the handling of AI risks across the EU membership, establishing a regional AI regulatory regime and regional AI institutions to guide and oversee the implementation of the legislation in cooperation with national authorities. However, regional trade accords and AI governance regimes do not preclude inter-regional misalignment. For example, the EU AI Act builds in extraterritoriality in that it applies for foreign suppliers to the EU market. But, this does not preclude the possible emergence of regulatory inconsistency with other regimes that those same suppliers may face. At the same time, even less binding commitments such as joint communiqués issued by the G20 or G7 may promote certain types of aligned AI policy action among members and via taskings for other international bodies. Follow-up may take place through reporting processes, soft power means such as subsequent peer review and other methods (for example, even moral suasion among participating leaders). Success in improving regulatory cooperation (see Box 3) can yield tangible economic results. The OECD (2021, 19) reviewed evidence on the performance of international regulatory cooperation in various fields and found that such efforts delivered improvements in three areas in particular: regulatory effectiveness (especially for challenges that traverse international boundaries); economic efficiency (for example, limiting undue trade friction); and administrative efficiency (for example, domestic regulators may benefit from international intelligence and insights). In light of this past experience including aspects of the digital economy, it may well be that improved cooperation in the area of AI would yield similar types of results. What Is to Be Done? Recommendations In view of the pace of technological change in the AI sector, this paper has examined the concentration of AI innovation and considered factors in the regulatory environment that may support or hinder its responsible diffusion. Drawing on the combined findings from the three analytical sections, the paper recommends: →Designation of an international institution to lead on AI regulatory cooperation: Replacing the existing fragmented approach, national governments — perhaps via a joint decision of the G20 — may wish to establish a designated AI body at the global level to monitor regulatory Box 2: Regulatory Misalignment Can Be Costly The OECD (2021, 22) presents a taxonomy of trade costs for producers and traders operating internationally in cases of regulatory divergence. These include information costs related to obtaining and processing information on regulatory requirements (whereby more opaque and complex systems entail higher costs); specification costs to adjust products and services to different requirements (with potential reductions in economies of scale, higher labour and other input costs); conformity assessment costs (demonstrating compliance with different requirements may require additional costly tests, certification, inspections, audits); and other costs, including to the governments administrating the regime. The burdens of misalignment do not fall equally. Businesses in small economies could be at a particular disadvantage with respect to accessing models, software applications and quality data. Small businesses in those economies could be disproportionately affected, as they often lack capacity to track details of regulatory developments in foreign markets. In addition to misalignment, in some economies, the turbulence of changing regulation may create uncertainty that undermines incentives to invest. For larger economies such as China, the European Union and the United States, there may be sufficient scale in the domestic market to continue to fuel AI development even in a fragmented global AI economy. 39AI Innovation Concentration and the Governance Challenge itself provides for penalties or in the most serious cases withdrawal of use of generative AI services. A team from HSBC Global Research recently considered the impact of Chinese regulation on the development of generative AI in China (Liu et al. 2023). The authors note that the registration system for algorithms was operational and that a first batch of 12 LLMs had reportedly been approved by regulators in August 2023. All 12 originated with Chinese developers, including Alibaba’s Tongyi Qianwen, Baidu’s Ernie Bot and Tencent’s Hunyuan. The HSBC analysts note that these firms will have a number of options for profitable commercialization of their AI capabilities, such as AI computing power (capitalizing on large graphics processing unit inventories to supply computing services to clients); AI generated content; and model-as-aservice ( making models available or customizing models for customers). While American and other foreign firms have a lead in such commercialization globally, Chinese firms such as Baidu, Alibaba and Tencent (the so-called BAT) are moving to catch-up and will have a regulatory advantage in their domestic market. HSBC estimates that the generative AI market in China could reach a scale of US$10 billion by 2026 (Liu et al. 2023). The European Union The EU authorities are positively inclined toward AI innovation and use, subject to compliance with existing regulation and pending updates meant to protect users and ensure a contestable market. The EU rulebook is set for a significant upgrade. On December 8, 2023, the European Commission, Council and Parliament agreed on the terms for a landmark comprehensive law on AI, the EU Artificial Intelligence Act. The legal text is now being finalized and will be submitted to the European Council and Parliament for formal adoption in the coming months (Council of the European Union 2023). Once adopted, the official text of the EU AI Act will be published. It enters into force 20 days later. The application of the provisions will roll out in phases over the next three years. The prohibition on models posing unacceptable risks will apply six months after the law’s entry into force. Firms developing general purpose AI models will need to comply within 12 months. The European Commission will have 18 months to provide practical guidelines for the classification of high-risk AI systems. All AI Act rules will be applicable within 24 months in most cases, although operators of certain special purpose AI systems (for example, critical infrastructure; migration, asylum and border control management) will have an additional year to comply. The EU AI Act governance framework uses a technology-neutral, risk-based approach targeting a range of use cases (for example, see, Dentons 2023; Hoffmann 2023). The new legislation will complement other AI-relevant elements already in place as part of the European Union’s approach to governance in the digital sphere. And, it has an element of extraterritoriality: it will apply to providers serving the EU market, regardless of where they are based. The EU AI Act defines AI systems broadly, drawing on work from the OECD (Bamberg et al. 2024). As of the time of writing, the proposed text of the EU AI Act defined the scope of the legislation as being AI systems that are machine-based and “designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.”58 The act further elaborates that it is not meant to cover traditional software programs based on rules designed by natural persons to execute specific operations and lacking the capacity to infer.59 The pending AI Act employs a four-tiered risk structure (Madiega 2023; Hoffmann 2023; Dentons 2023; Heikkilä 2023; Bamberg et al. 2024), with special supplemental provisions for general purpose AI models. → The AI Act would explicitly ban AI practices that pose an unacceptable risk to people’s safety, livelihoods and rights. According to the pending legislation, this would include AI systems that deploy: – harmful cognitive behavioural manipulation techniques; – emotional recognition in the workplace or educational institutions (except for medical and safety reasons); 58 This text draws on language from the marked-up legislative text of the EU AI Act, art. 3(1), which was accessed via Bamberg et al. (2024). 59 This text draws on language from the marked-up legislative text of the EU AI Act, recital 6, and accessed via Bamberg et al. (2024). 40 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt – algorithms to exploit specific vulnerable groups, for example, those with physical or mental disability) and thereby cause harm; – algorithms used by public authorities, or on their behalf, for social-scoring purposes and thereby causing detrimental or unfavourable treatment; – untargeted scraping of facial images from the internet or closed-circuit television footage to create facial recognition databases; – real-time, remote biometric identification in publicly accessible spaces for law enforcement (albeit with exceptions for certain grievous and urgent matters such as threat of terrorist attack); – biometric categorization to infer sensitive data such as sexual orientation or religious beliefs; and – predictive policing of individuals based solely on profiling or personality traits. → The AI Act would tightly regulate highrisk AI systems that could adversely affect safety or fundamental rights. Providers and deployers of such systems will face detailed and comprehensive obligations. Among other requirements, such systems would be subject to governance and technical stipulations covering such areas as transparency, risk management, accountability, data governance, accuracy, robustness and cybersecurity. Such high-risk AI systems will need to be subject to human oversight and control. Data for training and use in these AI systems will need to be high quality, relevant and unbiased. In light of the risks, the European Union will require registration of such AI systems in a special database. High-risk systems will be further categorized into two main groups, Annex II and Annex III systems (the latter will have an extra year to reach compliance, 36 months in total): – Annex II systems are used as a safety component of a product that is itself riskprone, covered by EU law and subject of a required conformity assessment. – Annex III systems serve various specific purposes: non-banned biometrics; critical infrastructure; educational and vocational training including systems related to program access, evaluation and monitoring of testing; human resource management in employment; access and use of essential public and private services; law enforcement; migrations, asylum and border control management; administration of justice and democratic processes. → Limited-risk AI systems such as chatbots and generative AI systems would be more lightly regulated but still be subject to a set of transparency obligations. Users will need to be informed of their interactions with such an AI system and the limitations of such systems. AIgenerated content would need to be flagged. → Lowand minimal-risk AI systems (for example, spam filters or video games) could be developed and used freely in the European Union, subject to other applicable laws. However, the EU AI Act envisages creation of codes of conduct for providers of limitedand low-risk systems to guide them toward voluntary compliance with higher AI standards. General-purpose AI models (foundation models, LLMs) will be subject to further obligations over their life cycle (Bamberg et al. 2024). These include maintaining current technical documentation of the model (including training and testing process, with evaluation results); transparency on properties of the model for downstream system providers; development of a policy for compliance with copyright law; and a detailed summary of the training data employed in the model’s development. In cases of systemic risk, generalpurpose AI model developers must implement risk assessment and mitigation measures, as well as incident response and reporting procedures. In addition, the EU AI Act establishes an institutional framework to support and enforce the operation of the legislation. Each member state must designate at least one national competent authority to supervise the AI Act implementation and track AI market developments. A new European Artificial Intelligence Board, with one national representative per member, will oversee this policy area at the EU level. A new EU AI Office within the European Commission is tasked with oversight of the general purpose AI models regime, supporting national authorities, and certain aspects of enforcement. A scientific panel of independent experts will support the EU institutions in this work. 41AI Innovation Concentration and the Governance Challenge With respect to enforcement, the legislation specifies substantial penalties that depend on the severity of the infringement. These vary between 1.5 percent and 7.0 percent of annual turnover for a firm, subject to certain minimum and maximum thresholds. As for other AI-relevant EU legislation that would operate in conjunction with the AI Act, a few illustrative examples are cited below: → Given the expansive exploitation of data through AI technology, the EU GDPR will be quite relevant. The GDPR empowers data subjects to exert some control over the use of their personal data via consent requirements and requires some privacy protections, particularly in relation to profiling and decision making. As with any EU business, AI providers must adhere to principles such as data minimization, purpose limitation and storage limitation when collecting and processing personal data. On the other hand, the GDPR does include allowances for data use for statistical and scientific research purposes. A European Parliament study found that the GDPR “can be interpreted and applied in such a way that it does not hinder beneficial application of AI to personal data, and that it does not place EU companies at a disadvantage in comparison with non-European competitors” (though a variety of specific issues still require clarification) (Sartor and Lagioia 2020, 79–80). GDPR adequacy determinations provide a degree of international access and portability with respect to EU data for certain firms based in a list of registered partner economies. As of January 15, 2024, 15 economies are covered by these decisions including leading AI host countries such as Canada, Israel, Japan, South Korea, the United Kingdom and the United States.60 → The EU Cybersecurity Act established a cybersecurity framework for products and services, operating under the EU Agency for Cybersecurity, aiming to harmonize EU-wide cybersecurity certification for information technology products, services and processes. As with other digital systems, AI systems would be covered. → The Digital Services Act (2022) imposes tiered obligations for online marketplaces and search 60 See https://commission.europa.eu/law/law-topic/data-protection/ international-dimension-data-protection/adequacy-decisions_en. engines, as well as protections for minors. Obligations include, among others, transparency in areas such as algorithmic decision making and recommendations, and for very large providers, due diligence with respect to annual assessment of systemic risks and regulatory access to databases, algorithms and premises (Beck and Worm 2023). → If an AI foundational model platform61 were to attain sufficient scale, for example, the Digital Market Act (2022) provisions might be invoked with consequences for AI system providers. Requirements could include third-party interoperability, limits on the use of personal data and constraints on prioritizing rankings of own products or services, among other possibilities that aim to restore fairness and contestability in the market (Yasar et al. 2023). → The Digital Governance Act (2022), which aims to facilitate access to publicly held data while maintaining privacy and confidentiality protections.62 → The Unfair Commercial Practices Directive (2005), as updated by subsequent guidance.63 The directive and guidance govern businessto-consumer transactions and representations made by vendors and other stakeholders (for example, providers of comparison tools). The guidance addresses (with explicit references to AI) such issues as tracking and targeting technologies and algorithmic personalization, among other risks. The United States As in other countries, AI in the United States is the subject of an emerging regulatory framework. This is not to say that it has been unregulated. At the federal level, the US government included some measures to define AI or address specific concerns such as AI applications in defence or transportation systems. A defence appropriations 61 AI foundational models are large-scale and intended to be used as a platform to which developers could add additional functionality. 62 See https://digital-strategy.ec.europa.eu/en/policies/data-governanceact-explained. 63 EC, Guidance on the interpretation and application of Directive 2005/29/EC of the European Parliament and of the Council concerning unfair business-to-consumer commercial practices in the internal market (see “Digital sector” at §4.2), [2021] OJ C 256/1; online: https://eur-lex.europa.eu/legal-content/EN/TXT/ PDF/?uri=CELEX:52021XC1229(05). 42 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt act in 2020 legislated the creation of the National Artificial Intelligence Initiative Office in the White House to oversee the US national AI strategy. (A number of US states have also legislated measures pertaining to AI, as noted in DPA [2023]). Operations in the AI sector have also been subject to prior existing regulatory provisions applicable more generally in the economy. Existing domestic law and regulation continued to apply, even in the AI sector. For example, a recent Congressional Research Service review of copyright protection in the sector found that existing copyright law may prove adequate and that additional experience with court challenges would be required to highlight any specific changes in the legal framework that might be required with respect to generative AI (Zirpoli 2023). As the US Federal Trade Commission Chair Khan (2023) put it in another context, “There is no AI exemption to the laws on the books.” Yet, the US approach to AI regulation has been gradually changing in recent years, starting with hardware. In the competition between China and the United States for AI leadership, the United States has focused on perceived security risks, moving to more tightly limit access to advanced AI-related technology of American origin. By means of an Executive Order dated September 15, 2022 (Biden 2022), President Biden expressly mandated that the Committee on Foreign Investment in the United States in conducting investment reviews should consider supply chain resilience and security with respect to AI and related areas of micro-electronics, among other areas. Then, on October 7, 2022 (Bureau of Industry and Security [BIS] 2022), the United States imposed controls on exports to China of certain advanced computing chips and equipment used to manufacture semiconductors, including certain foreign-produced items from third countries incorporating such advanced US technologies. The public notice (ibid., 1) justified the restrictions on national security and human rights grounds and noted the potential military applications for the covered items, as well as uses related to intelligence and security services. These steps went handin-hand with the CHIPS and Science Act passed earlier in 2022 by the US Congress, which delivered tremendous support for the US semiconductor sector (US$52.7 billion for manufacturing incentives, R&D, workforce development and more) and explicitly referenced its support for US leadership in the AI sector as an objective (White House 2022b). Also in October 2022, the White House released its “Blueprint for an AI Bill of Rights” (White House 2022a). The document set out five foundational principles for use in development of a framework for AI governance. These principles are meant to guide the design, use and deployment of automated systems. They cover protection from unsafe and ineffective systems, protection from algorithmic discrimination, protection from abusive data practices (while ensuring users have agency over data uses), user notice and explanation of AI use and impacts, and the ability to opt out and have access to a human (for example, to remedy problems). In order to get a better view of developments and concerns with respect to AI technology, members of the US administration held structured discussions with AI stakeholders during the first half of 2023. Drawing on this input and cross-referencing the “Blueprint for an AI Bill of Rights” and other recent administrative actions concerning AI, the White House proceeded to secure voluntary commitments from seven leading AI firms to ensure safety, security and trust in developing AI (White House 2023b). Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI signed on as of July21, 2023. In September 2023, eight further firms signed on: Adobe, Cohere, IBM, Nvidia, Palantir, Salesforce, Scale AI and Stability (Mitchell 2023). The key features of the eight US voluntary commitments (White House 2023b) are: → internal and independent external security testing of their AI systems before their release; → sharing information across the industry and with governments, civil society and academia on managing AI risks (for example, best practices for safety); → investing in cybersecurity and insider threat safeguards to protect proprietary and unreleased model weights; → facilitating ongoing third-party discovery and reporting of vulnerabilities in their AI systems; → developing robust technical mechanisms to ensure that users know when content is AI generated, such as a watermarking system; → publicly reporting their AI systems’ capabilities, limitations and areas of appropriate and inappropriate use; 43AI Innovation Concentration and the Governance Challenge → prioritizing research on the societal risks that AI systems can pose, including on avoiding harmful bias and discrimination, and protecting privacy; and → agreeing to develop and deploy advanced AI systems to help address society’s greatest challenges (for example, cancer prevention and mitigating climate change). The United States announced a further tightening of export controls on October 17, 2023 (BIS 2023). Restrictions were expanded to cover a greater range of advanced computing semiconductors that could potentially “enable the development and production of technologies such as artificial intelligence (AI) used in military applications.” The public notice states (ibid., 2), “These controls were strategically crafted to address, among other concerns, the PRC’s [People’s Republic of China’s] efforts to obtain semiconductor manufacturing equipment essential to producing advanced integrated circuits needed for the next generation of advanced weapon systems, as well as high-end advanced computing semiconductors necessary to enable the development and production of technologies such as artificial intelligence (AI) used in military applications.” The restrictions concern the 22 countries to which the United States maintains an arms embargo, including China. As noted by HSBC analyst Frank He (2023, 1), “The affected product coverage is wider than market expectations, which may cause a material impact on China’s AI infrastructure supply chain as we estimate that the chips that are newly added to the control list currently support over 90% of AI training workloads in China.” The immediate disruption will be limited as Chinese importers anticipated potential disruption and had built up inventories of requisite chips. Also, it is notable that Chinese domestic suppliers such as Huawei and Hygon have improved their chip offerings and can supply viable substitutes for some chips. While He (ibid.) recognizes that the US export controls will incentivize further development in the domestic sector, he argues, however, that a multiyear technology gap may persist in system-level engineering and sees that as a likely constraint on the AI sector in China. Areas of concern include, for example, chip-to-chip interconnection, software ecosystems and advanced node chip fabrications. The US government took a substantial step to strengthen its AI governance with the “Executive Order (EO) on Safe, Secure, and Trustworthy Artificial Intelligence” issued by President Biden on October 30, 2023 (Biden 2023; White House 2023c). The aim was to advance a coordinated, federal government-wide approach to position the United States for responsible leadership in AI development while addressing AI risks to safety and security. The framework seeks to capitalize and promote benefits from development of AI, while also taking care to mitigate harms. In view of partisan blockage in the US Congress, the Executive Order exploits authorities available to the president under existing law. Still, this action is relatively expansive considering the constraints. The Executive Order directs that executive agencies use their current powers to develop standards, tools and tests to ensure that AI systems are safe, secure and trustworthy. The approach is risk based, with special reporting measures included covering developers of powerful foundation models. It takes steps to address risks concerning privacy, workplace and other discrimination, and labour market dislocations. The initiative includes measures to support responsible AI use in health care and education. The Executive Order aims to boost AI-related human capital by supporting AI research, facilitating relevant skilled immigration and expansion of federal staffing of AI professionals. Administrative measures would facilitate ramping up use of AI in federal agencies. Openness would be cultivated via support for small business, international collaboration and deployment of AI in meeting challenges global challenges. The document emphasizes the goal of developing and implementing international standards for AI. On November 21, 2023, the US Federal Trade Commission (FTC) (2023) took a step moving further beyond the early voluntary approaches to AI governance. Under the FTC Act and other laws, the FTC is concerned with fraud, deception, infringements on privacy, unfair commercial practices and competition issues, including instances involving AI. Consequently, the FTC authorized the use of a compulsory process64 in investigations related to products and services produced or used in connection with AI. Recipients 64 The term “compulsory process” refers to various types of information or document requests — including but not limited to subpoenas, civil investigative demands, and orders for special reports — where the recipient of the request may be compelled to comply with the request by order of court. 44 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt of requests for information or documents may be legally compelled to comply by a court order. As of November 2023, the US Administration’s Office of Management and Budget has published for public consultation a draft memorandum with further concrete steps to promote AI governance, innovation and risk management in agencies of the federal government (Young 2023). With the exception of national security agencies, the memorandum would require each agency to designate a senior-level chief AI officer, develop an AI compliance plan corresponding to the memorandum, inventory the agency’s AI systems and risks (emphasizing safety and rights impacts), identify potential responsible uses of AI and remove barriers to delivering those systems. Non-compliant AI systems are to be terminated by August 1, 2024. The memorandum identifies specific issues for consideration with respect to each mandate, with milestones, standards, and relatively tight timelines for completion (measured in days and months for initial actions and then for ongoing requirements, depending on the action, annually or biennially). OECD Recommendation of the Council on Artificial Intelligence The OECD principles on AI within the OECD’s “Recommendation of the Council on Artificial Intelligence”65 state that: → AI should benefit people and the planet by driving inclusive growth, sustainable development and well-being. → AI systems should be designed in a way that respects the rule of law, human rights, democratic values and diversity, and they should include appropriate safeguards — for example, enabling human intervention where necessary— to ensure a fair and just society. → There should be transparency and responsible disclosure around AI systems to ensure that people understand when they are engaging with them and can challenge outcomes. 65 See OECD (2019). Also note that as of November 3, 2023, the OECD Legal Instruments online database reports that adherence to the recommendation now covers all 38 OECD member countries, plus eight others: Argentina, Brazil, Egypt, Malta, Peru, Romania, Singapore and Ukraine. → AI systems must function in a robust, secure and safe way throughout their lifetimes, and potential risks should be continually assessed and managed. → Organizations and individuals developing, deploying or operating AI systems should be held accountable for their proper functioning in line with the above principles. The OECD (2019) recommends that governments: → facilitate public and private investment in R&D to spur innovation in trustworthy AI; → foster accessible AI ecosystems with digital infrastructure and technologies, and mechanisms to share data and knowledge; → create a policy environment that will open the way to deployment of trustworthy AI system; → equip people with the skills for AI and support workers to ensure a fair transition; and → cooperate across borders and sectors to share information, develop standards and work towards responsible stewardship of AI. 45AI Innovation Concentration and the Governance Challenge Works Cited Aaronson, Susan Ariel. 2024. Data Disquiet: Concerns about the Governance of Data for Generative AI. CIGI Paper No. 290. Waterloo, ON: CIGI. www.cigionline.org/publications/data-disquiet-concernsabout-the-governance-of-data-for-generative-ai/. Aaronson, Susan Ariel and Patrick Leblond. 2018. “Another Digital Divide: The Rise of Data Realms and its Implications for the WTO.” Journal of International Economic Law 21(2): 245–72. https://doi.org/10.1093/jiel/jgy019. Bamberg, David, Martin Pflueger, Nicole Saurin, Stefan Schuppert, Jasper Siems, Dan Whitehead, Eduardo Ustaran and Leopold von Gerlach. 2024. “The EU AI Act: an impact analysis (part 1).” Engage, February 1. Hogan Lovells. www.engage.hoganlovells.com/knowledgeservices/ news/the-eu-ai-act-an-impact-analysis-part-1/. Beck, Benjamin and Ulrich Worm. 2023. “EU Digital Services Act’s Effects on Algorithmic Transparency and Accountability.” Mayer Brown, March 27. www.mayerbrown.com/en/perspectivesevents/publications/2023/03/eu-digital-services-actseffects-on-algorithmic-transparency-and-accountability. Bender, Emily M., Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In FAcct 2021: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–23. New York, NY: Association for Computing Machinery. https://doi.org/10.1145/3442188.3445922. Biden, Joseph R. 2022. “Ensuring Robust Consideration of Evolving National Security Risks by the Committee on Foreign Investment in the United States.” Executive Order 14083, September 15. Presidential Document. Federal Register (87) 181, September 20. www.govinfo.gov/ content/pkg/FR-2022-09-20/pdf/2022-20450.pdf. ———. 2023. “Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence.” The White House Briefing Room, October 30. www.whitehouse.gov/ briefing-room/presidential-actions/2023/10/30/ executive-order-on-the-safe-secure-and-trustworthydevelopment-and-use-of-artificial-intelligence/. BIS. 2022. “Commerce Implements New Export Controls on Advanced Computing and Semiconductor Manufacturing Items to the People’s Republic of China (PRC).” BIS, US Department of Commerce. October 7. www.bis.doc.gov/index.php/ documents/about-bis/newsroom/press-releases/31582022-10-07-bis-press-release-advanced-computingand-semiconductor-manufacturing-controls-final/file. ———. 2023. “Commerce Strengthens Restrictions on Advanced Computing Semiconductors, Semiconductor Manufacturing Equipment, and Supercomputing Items to Countries of Concern.” BIS, US Department of Commerce. October17. www.bis.doc.gov/index.php/documents/ about-bis/newsroom/press-releases/3355-2023-1017-bis-press-release-acs-and-sme-rules-final-js/file. Branstetter, Lee, C. Fritz Foley and Kamal Saggi. 2010. “Has the Shift to Stronger Intellectual Property Rights Promoted Technology Transfer, FDI, and Industrial Development?” The WIPO Journal 2 (1): 93–98. https://tind.wipo.int/record/28665. Breaux, Cory and Emin Dinlersoz. 2023. “How Many U.S. Businesses Use Artificial Intelligence?” US Census Bureau. November 28. www.census.gov/library/stories/2023/11/businessesuse-ai.html?utm_campaign=20231128msacos1ccstors&utm_ medium=email&utm_source=govdelivery. Bremmer, Ian and Mustafa Suleyman. 2023. “The AI Power Paradox: Can States Learn to Govern Artificial Intelligence — Before It’s Too Late?” Foreign Affairs 102 (5): 26–43. www.foreignaffairs.com/world/artificial-intelligencepower-paradox. Buildd. 2022. “ByteDance — The world’s most valuable startup!” April 13. Bengaluru, India: Buildd.co. https://buildd.co/ product/bytedance-worlds-most-valuable-startup. CAC. 2021. “Provisions on the Management of Algorithmic Recommendations in Internet Information Services.” CAC, Ministry of Industry and Information Technology, Ministry of Public Safety, State Administration of Market Regulation. Translation by China Law Translate, 2022. www.chinalawtranslate.com/en/algorithms/. ———. 2022. “Provisions on the Administration of Deep Synthesis Internet Information Services.” CAC, Ministry of Industry and Information Technology, Ministry of Public Safety. Translation by China Law Translate, 2022. www.chinalawtranslate.com/en/deep-synthesis/. ———. 2023. “Measures for the Management of Generative Artificial Intelligence Services (Draft for Comment).” Cyberspace Administration of China (CAC), Ministry of Science and Technology, and the Ministry of Public Security. Translation by Seaton Huang, Helen Toner, Zac Haluza, Rogier Creemers and Graham Webster, at DigiChina, Stanford University, 2023. https://digichina.stanford.edu/work/translationmeasures-for-the-management-of-generative-artificialintelligence-services-draft-for-comment-april-2023/. Cavazos Cepeda, Ricardo H. and Douglas C. Lippoldt. 2010. “The Strengthening of IPR Protection: Policy Complements.” The WIPO Journal 2 (1): 99–122. https://tind.wipo.int/record/28665. 46 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt CB Insights. 2023a. “Global Unicorn Club.” Online database. October 19. www.cbinsights.com/research/ report/unicorn-startups-valuations-headcountinvestors/?utm_source=CB+Insights+Newsletter&utm_ campaign=bfc39b1e66-newsletter_general_ thurs_2023_10_26&utm_medium=email&utm_ term=0_9dc0513989-bfc39b1e66-90194953. ———. 2023b. “The open-source AI development market map.” Research brief, CB Insights Research. October 13. www.cbinsights.com/research/ open-source-ai-development-market-map/. Chen, Hailin, Fangkai Jiao, Xingxuan Li, Chengwei Qin, Mathieu Ravaut, Ruochen Zhao, Caiming Xiong and Shafiq Joty. 2023. “ChatGPT’s One-year Anniversary: Are Open-Source Large Language Models Catching up?” arXiv, November29. https://doi.org/10.48550/arXiv.2311.16989. Christensen, Clayton M. 2016. The Innovator’s Dilemma: When New Technologies Cause Great Firms to Fail. Boston, MA: Harvard Business Review Press. Chui, Michael, Bryce Hall, Helen Mayhew, Alex Singla and Alex Sukharevsky. 2022. “The state of AI in 2022 — and a half decade in review.” McKinsey & Company. December 6. www.mckinsey.com/capabilities/quantumblack/our-insights/ the-state-of-ai-in-2022-and-a-half-decade-in-review#/. Ciuriak, Dan and Anna Artyushina. 2023. “Canada Needs an Artificial Intelligence Agency.” Opinion, Centre for International Governance Innovation, October 5. www.cigionline.org/ articles/canada-needs-an-artificial-intelligence-agency/. Council of the European Union. 2023. “Artificial intelligence act: Council and Parliament strike a deal on the first rules for AI in the world.” Press release, December 9. www.consilium.europa.eu/en/press/pressreleases/2023/12/09/artificial-intelligence-act-council-andparliament-strike-a-deal-on-the-first-worldwide-rules-for-ai/. Dentons. 2023. “The New EU AI Act – the 10 key things you need to know now.” December 14. www.dentons.com/ en/insights/articles/2023/december/14/the-neweu-ai-act-the-10-key-things-you-need-to-know-now. DPA. 2023. “Activity Tracker: ML and AI Development.” DPA website. Viewed on November 18, 2023. https://digitalpolicyalert.org/activity-tracker?offset=0&l imit=10&activity=9&period=2019-12-31,2023-11-17. DSIT. 2023. “Emerging Processes for Frontier AI Safety.” Conference Background Paper, AI Safety Summit. DSIT, United Kingdom, October. https://assets.publishing. service.gov.uk/media/653aabbd80884d000df71bdc/ emerging-processes-frontier-ai-safety.pdf. Ederer, Florian and Bruno Pellegrino. 2023. “The Great Startup Sellout and the Rise of Oligopoly.” AEA Papers and Proceedings 113 (May): 274–78. https://papers.ssrn.com/sol3/papers. cfm?abstract_id=4322834. Epic Games. 2023. “Introducing the NoAI and CreatedWithAI tags.” Customer Help, February 20. https://marketplacehelp.epicgames.com/s/ article/Introducing-the-NoAI-andCreatedWithAI-tags?language=en_US. European Parliament. 2024. “Artificial Intelligence Act: MEPs adopt landmark law.” Press release, March 13. www.europarl.europa.eu/news/en/ press-room/20240308IPR19015/artificialintelligence-act-meps-adopt-landmark-law. Fathi, Eli and Peter MacKinnon. 2023. “AI Guardrails Are Urgently Needed, but That’s Just a Start.” Opinion, Centre for International Governance Innovation, October 11. www.cigionline.org/articles/ai-guardrailsare-urgently-needed-but-thats-just-a-start/. Fay, Robert and Rohinton Medhora. 2021. “A Global Governance Framework for Digital Technologies.” T20 Italy 2021, Task Force 4, Policy Brief 5. www.t20italy.org/2021/08/26/aglobal-governance-framework-for-digital-technologies/. Ferguson, Christopher and Julie He. 2023. “China’s New Rules For Generative AI: An Emerging Regulatory Framework.” Fasken Technology, Media and Telecommunications Bulletin. August 29. www.fasken.com/en/knowledge/2023/08/ chinas-new-rules-for-generative-ai#authors. FTC. 2023. “FTC Authorizes Compulsory Process for AI-related Products and Services.” Press release, November 21. www.ftc.gov/news-events/news/press-releases/2023/11/ ftc-authorizes-compulsory-process-ai-related-products-services. Fujian. 2021. “Outline of the 14th Five-Year Plan (2021-2025) for National Economic and Social Development and Vision 2035 of the People’s Republic of China.” The People’s Government of Fujian Province. September 8. www.fujian.gov.cn/ english/news/202108/t20210809_5665713.htm. GOV.UK. 2023. “The Bletchley Declaration by Countries Attending the AI Safety Summit. November 1–2.” Policy Paper. November 1. www.gov.uk/government/publications/ ai-safety-summit-2023-the-bletchley-declaration/ the-bletchley-declaration-by-countries-attendingthe-ai-safety-summit-1-2-november-2023. Grassano, Nicola, Héctor Hernandez Guevara, Péter Fako, Elisabeth Nindl, Aliki Georgakaki, Ela Ince, Lorenzo Napolitano, Francesco Rentocchini and Alexander 47AI Innovation Concentration and the Governance Challenge Tuebke. 2022. “The 2022 EU Industrial R&D Investment Scoreboard.” Luxembourg: Publications Office of the European Union. https://iri.jrc.ec.europa.eu/ scoreboard/2022-eu-industrial-rd-investment-scoreboard. Hatzius, Jan, Dominic Wilson, Joseph Briggs, Vickie Chang, Devesh Kodnani and Giovanni Pierdomenico. 2023. “Macro Outlook 2024: The Hard Part is Over.” (Redacted version.) Goldman Sachs. November 8. www.goldmansachs.com/ intelligence/pages/gs-research/macro-outlook2024-the-hard-part-is-over/report.pdf. He, Frank. 2023. “China Technology — AI Infrastructure: Looking for Alternatives.” Equity Research Report. HSBC Global Research. October 19. Heikkilä, Melissa. 2023. “Five things you need to know about the EU’s new AI Act.” MIT Technology Review, December 11. www.technologyreview.com/2023/12/11/1084942/ five-things-you-need-to-know-about-the-eus-new-ai-act/. Ho, Lewis, Joslyn Barnhart, Robert Trager, Yoshua Bengio, Miles Brundage, Allison Carnegie, Rumman Chowdhury et al. 2023. “International Institutions for Advanced AI.” arXiv, July 11. https://arxiv.org/abs/2307.04699. Hoffmann, Mia, 2023. “The EU AI Act: A Primer.” CSET Blog, September 26. https://cset.georgetown.edu/ article/the-eu-ai-act-a-primer/ . IFAC-BIAC. 2018. “Regulatory Divergence: Costs, Risks, Impacts.” www.ifac.org/_flysystem/azure-private/publications/ files/IFAC-OECD-Regulatory-Divergence.pdf. Innovation, Science and Economic Development. 2023. “The Artificial Intelligence and Data Act (AIDA) – Companion document.” Government of Canada. March 13. https://ised-isde.canada.ca/site/innovation-better-canada/ en/artificial-intelligence-and-data-act-aidacompanion-document. Jones, Elliot. 2023. “Explainer: What is a foundation model?” Ada Lovelace Institute. July 17. www.adalovelaceinstitute.org/ resource/foundation-models-explainer. Kaye, Kate. 2022. “Will nationalism end the golden age of global AI collaboration?” Protocol, November 1. https://web.archive.org/web/20221101100423/www. protocol.com/enterprise/china-us-ai-open-source. Khan, Lina M. 2023. “Statement of Chair Lina M. Khan Regarding the Joint Interagency Statement on AI.” FTC, April 23. www.ftc.gov/system/files/ftc_gov/pdf/remarks-of-chairlina-m-khan-re-joint-interagency-statement-on-ai.pdf. Kiriyama, Nobuo. 2012. “Trade and Innovation: Synthesis Report.” OECD Trade Policy Papers No. 135. Paris, France: OECD Publishing. http://dx.doi.org/10.1787/5k9gwprtbtxn-en. Krugman, Paul. 1995. Development, Geography and Economic Theory. Cambridge, MA: The MIT Press. Leblond, Patrick. 2021. “Uploading CPTPP and USMCA Provisions to the WTO’s Digital Trade Negotiations Poses Challenges for National Data Regulation: Example from Canada.” In Big Data and Global Trade Law, edited by Mira Burri, 301–15. Cambridge, UK: Cambridge University Press. http://dx.doi.org/10.1017/9781108919234.019. Lee, Kai-Fu. 2018. AI Superpowers: China, Silicon Valley and the New World Order. Boston, MA: Houghton Mifflin Harcourt. Lippoldt, Douglas and Mark. F. Schultz. 2014. “Uncovering Trade Secrets — An Empirical Assessment of Economic Implications of Protection for Undisclosed Data.” OECD Trade Policy Papers No. 167. Paris, France: OECD Publishing. http://dx.doi.org/10.1787/5jxzl5w3j3s6-en. Liu, Charlene, Ritchie Sun, Peishan Wang and Mark McDonald. 2023. “China Generative AI — Part One: A Guide to Monetisation.” HSBE Global Research. October 17. Luong, Ngor and Zachary Arnold. 2021. “China’s Artificial Intelligence Industry Alliance: Understanding China’s AI Strategy Through Industry Alliances.” Center for Security and Emerging Technology. Data Brief. May. https://cset.georgetown.edu/wp-content/uploads/CSETChinas-Artificial-Intelligence-Industry-Alliance-1.pdf. Madiega, Tambiama. 2023. “Artificial intelligence act.” Briefing. European Parliament, June. www.europarl.europa.eu/ RegData/etudes/BRIE/2021/698792/ EPRS_BRI(2021)698792_EN.pdf. Maguire, James. 2023. “100+ Top Artificial Intelligence (AI) Companies 2023.” eWeek, May 29. Maslej, Nestor, Loredana Fattorini, Erik Brynjolfsson, John Etchemendy, Katrina Ligett, Terah Lyons, James Manyika et al. 2023. Artificial Intelligence Index Report 2023. Stanford, CA: AI Index Steering Committee, Institute for HumanCentered AI. https://aiindex.stanford.edu/wp-content/ uploads/2023/04/HAI_AI-Index-Report_2023.pdf. McCarthy, John, Marvin Minsky, Nathaniel Rochester and Claude Shannon. 1955. “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.” Dartmouth College. August 31. www-formal.stanford.edu/jmc/history/ dartmouth.pdf. 48 CIGI Papers No. 292 — April 2024 • Douglas Lippoldt Mishra, Saurabh, Robert Koopman, Giuditta De Prato, Anand Rao, Israel Osorio-Rodarte, Julie Kim, Nikola Spatafora, Keith Strier and Andrea Zaccaria. 2023. “AI specialization for pathways of economic diversification.” Scientific Reports13(1). https://doi.org/10.1038/s41598-023-45723-x. Mitchell, Billy. 2023. “Eight more tech companies join White House commitment to manage AI risks.” Fedscoop, September 12. https://fedscoop.com/eight-more-tech-companiesjoin-white-house-commitment-to-manage-ai-risks/. Moore, Olivia. 2023. “How Are Consumers Using Generative AI?” Andreessen Horowitz, September 13. https://a16z.com/ how-are-consumers-using-generative-ai/. Murgia, Madhumita and John Thornill. 2023. “Superintelligent AI — can chatbots think?” November 27, in Tech Tonic, produced by the Financial Times, podcast, 30:23. www. ft.com/content/36af32cf-fa29-466d-bcff-8a8a898829d2. NIST. 2019. “U.S. Leadership in AI: A Plan for Federal Engagement in Developing Technical Standards and Related Tools.” www.nist.gov/system/files/documents/2019/08/10/ ai_standards_fedengagement_plan_9aug2019.pdf. ———. 2023. “Artificial Intelligence Risk Management Framework (AIRMF1.0).” January. https://doi.org/10.6028/NIST.AI.100-1. OECD. 2019. “Recommendation of the Council on Artificial Intelligence.” OECD Legal Instruments (OECD/LEGAL/0449). May 21. https://legalinstruments.oecd.org/ en/instruments/OECD-LEGAL-0449. ———. 2021. “International Regulatory Co-operation.” OECD Best Practice Principles for Regulatory Policy series. Paris, France: OECD. https://doi.org/10.1787/5b28b589-en. ———. 2024. “Explanatory Memorandum on the Updated OECD Definition of an AI System.” OECD Artificial Intelligence Papers No. 8. www.oecd-ilibrary.org/science-andtechnology/explanatory-memorandum-on-the-updatedoecd-definition-of-an-ai-system_623da898-en. OECD.AI. 2021. “Database of national AI Policies, powered by EC/OECD (2021).” OECD Artificial Intelligence Policy Observatory. Paris, France: OECD. https://oecd.ai/en/dashboards/overview. Office of Management and Budget. 2019. “Identifying Priority Access or Quality Improvements for Federal Data and Models for Artificial Intelligence Research and Development (R&D), and Testing; Request for Information.” Federal Register Document No. 2019-14618. July 10. www.federalregister.gov/ documents/2019/07/10/2019-14618/ identifying-priority-access-or-quality-improvementsfor-federal-data-and-models-for-artificial. Park, Walter G. and Douglas Lippoldt. 2005. “International Licensing and the Strengthening of Intellectual Property Rights in Developing Countries during the 1990s.” OECD Economic Studies 40: 7–48. https://doi.org/10.1787/ eco_studies-v2005-art2-en. ———. 2014. “Channels of technology transfer and intellectual property rights in developing countries.” In Intellectual Property for Economic Development, edited by Sanghoon Ahn, Bronwyn H. Hall and Keun Lee, 33–89. Cheltenham, UK: Edward Elgar. Raskin, Asa and Tristan Harris. 2023. “The Promise and Peril of Open Source AI with Elizabeth Seger and Jeffrey Ladish,” November 21, in Your Undivided Attention, produced by Center for Humane Technology, podcast, 38:44. https://your-undivided-attention.simplecast.com/ episodes/the-promise-and-peril-of-open-source-aiwith-elizabeth-seger-and-jeffrey-ladish-Kv4pNW3s. Roberts, Huw, Josh Cowls, Jessica Morley, Mariarosaria Taddeo, Vincent Wang and Luciano Floridi. 2021. “The Chinese approach to artificial intelligence: an analysis of policy, ethics, and regulation.” AI and Society 36: 59–77. https://doi.org/10.1007/s00146-020-00992-2. SAG-AFTRA. 2023a. “TV/Theatrical 2023: Regulating Artificial Intelligence.” Los Angeles, CA: SAG AFTRA. www.sagaftra.org/files/sa_documents/AI TVTH.pdf. ———. 2023b. “TV/Theatrical Contracts 2023: Summary of Tentative Agreement.” November 16. Los Angeles, CA: SAG-AFTRA. www.sagaftra.org/files/sa_documents/ TV-Theatrical_23_Summary_Agreement_Final.pdf. Sartor, Giovanni and Francesca Lagioia. 2020. “The impact of the General Data Protection Regulation (GDPR) on artificial intelligence.” European Parliamentary Research Service, June. www.europarl.europa.eu/RegData/etudes/ STUD/2020/641530/EPRS_STU(2020)641530_EN.pdf. Sheehan, Matt. 2023. “China’s AI Regulations and How They Get Made.” Carnegie Endowment for International Peace Working Paper. July. https://carnegieendowment.org/2023/07/10/ china-s-ai-regulations-and-how-they-get-made-pub-90117. Suominen, Kati. 2021. “The CPTPP’s Impacts on Digital Trade and the Path Forward.” Center for Strategic and International Studies. October. https://csis-website-prod.s3.amazonaws.com/ s3fs-public/publication/211027_Suominen_CPTPP_FullReport. pdf?VersionId=6TBKVom7OctkpyAb1NDJmR0m0wwWpIyx. Tsai, Charles, Eloi Laourou, Mattias Ganslandt and Douglas Lippoldt. 2011. “Enhancing Market Openness, Intellectual Property Rights, and Compliance Through Regulatory Reform in Israel.” Paris, France: OECD. www.oecd.org/israel/48262991.pdf.